feat: implement end-to-end runtime profiling and telemetry

[rapheljr]

Summary

Adds a zero-overhead runtime telemetry and end-to-end profiling subsystem to TurboFieldfare without changing inference behavior, tensor layouts, or kernel performance.

Features Included

• RuntimeTelemetry Core (RuntimeTelemetry.swift): Thread-safe metrics collector capturing:
  • End-to-end timing (TTFT, prefill tok/s, decode tok/s, total tokens)
  • CPU phase breakdown (token prep, scheduling, cache lookup, router handoff, SSD prep, sampling, GPU sync wait)
  • Expert cache efficiency (slots/layer, hits, misses, hit rate %, evictions, loads per layer & expert)
  • SSD I/O stats (read ops, total bytes, avg/min/max read latency, peak concurrency, effective MB/s)
  • Metal GPU interval timings (cb1, cb2, shared expert FFN, LM head)
  • System memory footprint (Process RSS, common weights, expert cache capacity, KV cache, prefill scratch)
• CLI Options (TurboFieldfareCLI):
  • --profile: Formatted ASCII text profiling summary table printed to stdout after generation.
  • --profile-json: Machine-readable JSON metrics report printed to stdout.
• Documentation (docs/PROFILING.md): Complete profiling architecture, CLI guide, schema definitions, and benchmarking methodology.
• Unit Tests (ProfilingTests.swift): 100% test coverage verifying metrics accumulation, thread safety, zero-overhead guards, and format outputs.

Verification

• ./Scripts/test.sh: Passed 14/14 unit tests cleanly.
• swift build -c release: Built release executables successfully.