fix(metrics): throughput in output tokens/s, measured per program

[YoanSallami]

Summary

• Throughput now means output tokens per second (reasoning included), instead of LM calls per second. Same for RewardThroughput and OptimizerThroughput. TokensPerSecond still counts input + output tokens. Throughput values recorded before this change are calls/s and are not comparable.
• The throughput denominator is measured per program. The per-phase wall-clock lived in one thread-local stack, so programs evaluated concurrently on one event loop (e.g. the cells of a tuner sweep) interleaved their op_scope enter/exit on it: time was credited to the wrong phase, and each program's calls were divided by a shared clock.
  • Each trainer now owns a PhaseClock; the trainer's inference / reward / optimizer scopes accrue to it, and the LM/EM operational metrics read the clock of the program they are bound to.
  • A phase accrues while any context has it as its innermost scope on that clock, so concurrent regions of one phase count once, and the optimizer phase still gets its self-time only.
  • op_scope(phase) without a clock keeps the thread-local default, so existing callers are unchanged.
• Version bump to 0.10.004.

Test plan

• op_scope_test.py: new tests for self-time nesting, two programs interleaving phases on one loop, concurrent regions of one phase, and a nested program on another clock
• lm_metrics_test.py updated for output tokens/s
• uv run pytest synalinks/src/metrics synalinks/src/backend synalinks/src/trainers synalinks/src/modules/language_models synalinks/src/modules/embedding_models: 550 passed
• ruff check on the changed files

🤖 Generated with Claude Code