feat(DAH-2254): surface GPU splitting in lium CLI/SDK

[fortunelucky777]

DAH-2254 — Surface GPU splitting in lium CLI/SDK

Summary

• ExecutorInfo Pydantic model gains min_gpu_count_for_rental: int | None and available_gpu_count: int | None. Both pre-existing on the backend payload — this PR just deserializes them.
• Client.ls(widen_for_splitting: bool = False) — opt-in widen. When True with gpu_count=N set, also returns splittable nodes where min_gpu_count_for_rental ≤ N ≤ available_gpu_count. Default OFF preserves SDK semver — third-party agent code that calls Lium().ls(gpu_count=1) keeps strict semantics.
• lium ls --count N (the user-facing CLI) passes widen_for_splitting=True so the marketplace UI and the CLI agree.
• @machine("XxY") decorator unchanged — it calls sdk.ls() with no kwargs at `decorators.py:49`, so the widen path never fires. A comment block at `decorators.py:43-48` documents this for future editors.
• Rich-table Config column gains ↯ from N suffix on splittable rows (mirrors the existing (DinD) suffix pattern).
• lium ls --format json exposes min_gpu_count_for_rental so AI-agent consumers can decide programmatically.
• Version bump 0.0.24 → 0.0.25, matched in pyproject.toml + lium/__about__.py (release workflow cross-checks).

Files

• lium/sdk/models.py — two new optional fields on ExecutorInfo
• lium/sdk/client.py — _is_splittable_for_count predicate helper (reciprocal-parity comment with frontend) + widen_for_splitting kwarg + payload deserialization in _dict_to_executor_info
• lium/sdk/decorators.py — clarifying comment above sdk.ls() call
• lium/cli/ls/actions.py — CLI passes widen_for_splitting=True
• lium/cli/ls/command.py — --count help text updated
• lium/cli/ls/display.py — Rich-table marker + JSON field
• pyproject.toml, lium/__about__.py — version bump
• test/fixtures/splittable_executors.json (new) — shared fixture (byte-identical with frontend repo)
• test/test_client_ls_widen.py (new) — 5 tests covering widen=True/False, default-off invariant, JSON field, Rich-table marker

Acceptance criteria

• CLI-AC feat(pod): DAH-1319 - improve pod run to deploy without docker hub credential #1 — ExecutorInfo.min_gpu_count_for_rental + available_gpu_count populated from API
• CLI-AC feat: DAH-1330 - add pods commands #2 — lium ls table marks splittable rows with `↯ from N`
• CLI-AC feat: DAH-1341 - search executors by location #3 — lium ls --format json includes min_gpu_count_for_rental
• CLI-AC feat: DAH-1353 - cache most used docker templates #4 — lium ls --count N widens to splittable; positive test covers widen=True/False
• CLI-AC Remove unused dependencies from pyproject.toml #5 — @machine decorator stays strict (no kwargs to sdk.ls(); default-off invariant asserted by inspect.signature test)
• CLI-AC fix botched refactor (it's lium.io) #6 — version matched: pyproject.toml:7 and lium/__about__.py:3 both at `0.0.25`

Test plan

• `pytest test/test_client_ls_widen.py` — 5/5 passing
• `python3 -c "import inspect; from lium.sdk.client import Client; assert inspect.signature(Client.ls).parameters['widen_for_splitting'].default is False"` — Gate D OK
• Source grep confirms `widen_for_splitting: bool = False` at `lium/sdk/client.py:506`
• Manual smoke: `lium ls` against staging, verify marker on a known splittable node; lium ls --count 1 returns native 1xGPU + splittable bigger nodes with min ≤ 1; lium ls --format json | jq '.[0].min_gpu_count_for_rental' returns the value.

Companion PR

Datura-ai/lium-io-frontend#204 — `feat/2254-gpu-splitting-filter` surfaces the same signal in the marketplace UI. The frontend's `matchesGpuCountFilter` predicate and this PR's `_is_splittable_for_count` predicate share a byte-identical JSON fixture; reciprocal code comments at `PodFiltersProvider.tsx` and `client.py` link them.

Source artifacts

• Spec: `.omc/specs/deep-interview-dah-2254-gpu-splitting.md` (deep-interview final ambiguity 12.5%)
• Plan: `.omc/plans/dah-2254-gpu-splitting.md` (Critic-APPROVED consensus, 14/14 ACs traced)

Notes for reviewer

• Client.ls() default OFF is deliberate. SDK consumers that call Lium().ls(gpu_count=1) programmatically (AI agents per spec) get the same result-set as before this PR. Only lium ls opts in; only the CLI passes widen_for_splitting=True.
• If you ever add gpu_count=... to the sdk.ls() call in decorators.py, decide explicitly about widen_for_splitting — the comment block flags this.
• The shared fixture path is test/fixtures/splittable_executors.json (mirrored at src/contexts/__fixtures__/splittable-executors.json in the frontend repo). Any change to one must be made byte-for-byte to the other; Gate C diff catches drift.