docs(cli): populate --help text for every command and flag across all CLI modules

[aarambh-darshan]

Summary

Before this PR, running aarambh-studio --help listed 7 of 10 subcommands
with blank descriptions (train, infer, eval, quantise, convert,
distill, finetune, selflearn), and most flags across the CLI showed no
description at all in their --help output. The only commands with help text
were agent (added in Phase 47) and serve, and the Phase 46 finetune rlaif
subcommand.

This PR adds human-readable help = "..." text and /// doc comments (which
clap renders as about) to every command struct and every flag across
all 9 CLI command modules, so <command> --help is now self-explanatory — a new
user can run any --help and understand what the command does and what each
flag means, without reading separate docs.

No behavior, flag names, types, defaults, or logic changed. This is a
documentation/UX-only change to the clap definitions.

What changed

All edits are under aarambh-studio/aarambh-studio/src/cmd/. Each file got a
/// doc comment on its command struct (rendered as the command about) and a
help = "..." string on every #[arg(...)]:

File	What was added
cmd/train.rs	TrainArgs about + --config help
cmd/infer.rs	InferArgs about + help on all ~40 flags (config, model, tokenizer, image/video/document/audio, pages, dpi, frames, frame-sampling, prompt, sampling, thinking, speculative, tools, safety, self-learn, forgetting, best-of-N, selection, ground-truth, …)
cmd/eval.rs	EvalArgs about + help on all ~20 flags (tasks, limit, kv-cache-report, forgetting-*, …)
cmd/quantise.rs	QuantiseArgs about + help on all 8 flags (format, output, calibration-samples, group-size, …)
cmd/convert.rs	ConvertArgs about + help on all 10 flags (from/to layouts, tokenizer, …)
cmd/serve.rs	help on all 15 flags (host, port, batch, prefill, api-key-env, cors-origin, …) — about was already present
cmd/distill.rs	DistillArgs about + help on every flag of TrainArgs/PrepareOfflineArgs/TrainOfflineArgs/EvaluateArgs
cmd/finetune.rs	FinetuneArgs about + doc comments on all 18 FinetuneCommand variants + help on every flag of FinetuneRunArgs/MergeArgs/GrpoArgs/DpoArgs/VlmFinetuneArgs (Phase 46 rlaif help left as-is)
cmd/selflearn.rs	SelflearnArgs about + doc comments on all 6 SelflearnCommand variants + help on every flag of StartArgs/SelflearnRunArgs/StatsArgs/ResetArgs/ForgettingReportArgs/ForgettingArgs

cmd/agent.rs was already complete (Phase 47) — verified and left untouched.
main.rs was not modified: clap already inherits each subcommand's about
from the wrapped Args struct's doc comment (the existing pattern), so
populating the Args struct doc comments in cmd/*.rs is sufficient to surface
every subcommand in the top-level --help list.

The verbatim --help reference snapshot at docs/cli-commands-raw-help.txt was
regenerated from the rebuilt binary (1130 → 1946 lines, reflecting the newly
populated descriptions).

Style followed (matches the existing agent command)

• One short, clear sentence per flag, no trailing period in help= (clap appends [default: ...] automatically — not duplicated).
• Enum-like flags list allowed values, e.g. Thinking budget: none, low, medium, high, or max.
• Repeatable flags note (repeatable).
• Phase references kept where relevant (e.g. best-of-n → "(Phase 45)").
• No TODO/FIXME/HACK/todo!()/unimplemented!() markers added (release-audit clean).

Proof it renders

Top-level --help (every subcommand now described; previously 7 were blank):

Commands:
  agent      Run a bounded caller-executed long-horizon tool-use chain
  train      Train a model checkpoint from a TOML training configuration
  infer      Generate text, multimodal, tool-use, speculative, best-of-N, or self-learning completions from a trained checkpoint
  eval       Evaluate a checkpoint on capability, forgetting, or KV-cache probes
  quantise   Quantise a trained checkpoint into a smaller GGUF format
  convert    Convert checkpoints between HF SafeTensors and GGUF, or expand vocabularies
  distill    Distil a smaller student model from a frozen local or dataset teacher
  finetune   Fine-tune adapters with SFT, DoRA, GRPO, DPO, RLAIF, or merge
  selflearn  Self-learning loop operator: start, flush, replay, stats, or reset
  serve      Start the local OpenAI-compatible inference server

train --help (was blank):

Train a model checkpoint from a TOML training configuration
Usage: aarambh-studio train --config <CONFIG>
Options:
      --config <CONFIG>  Training/model TOML configuration path

infer --help now describes every one of its ~40 flags, e.g.:

      --video <VIDEO>          Video file path for temporal video inference (mutually exclusive with --image)
      --thinking <THINKING>   Thinking budget: none, low, medium, high, or max [default: none]
      --best-of-n <BEST_OF_N> Generate N independent candidates and select the best (Phase 45)

Verification

• cargo fmt --all --check — clean
• cargo check -p aarambh-studio --all-targets --locked — passes
• cargo clippy -p aarambh-studio --all-targets --locked -- -D warnings -D clippy::undocumented_unsafe_blocks — zero warnings
• Rebuilt the debug binary and re-captured --help for every command and subcommand; previously-blank descriptions are now populated.
• No new dependencies; no flag names, types, defaults, or logic changed.

Scope/non-goals

• This PR is docs/UX only — it does not change any runtime behavior.
• The Phase 47 sandboxed-execution flags (--execute-tools, --allow-tool, --exec-*) already had help text and are unchanged here.
• The Phase 46 finetune rlaif help text was already complete and is unchanged.

Related

• Builds on Phase 47 (feat: Phase 47 — sandboxed tool execution) which established the help-text style in cmd/agent.rs.
• Reference snapshot: docs/cli-commands-raw-help.txt (regenerated for 4.0.0-alpha.7).