feat(tf2): implement training profiler support

## Summary

TF2 currently raises `NotImplementedError` when `training.profiling` or `training.enable_profiler` is configured, although both options are exposed by the common schema.

## Scope

- Define and implement the TF2 semantics of both profiler options.
- Own profiler start, step/range control, trace export, and cleanup through a dedicated lifecycle component.
- Honor the configured output file or directory.
- Keep profiler control outside compiled train-step functions.
- Prevent distributed ranks from writing colliding traces.
- Preserve a zero-overhead disabled path.

## Acceptance criteria

- Each profiler option works independently and in combination.
- Valid traces are written to the configured location.
- Resources are finalized on normal completion and exceptions.
- Distributed output follows a documented chief-only or rank-suffixed policy.
- Tests cover lifecycle ordering, trace creation, disabled behavior, and unchanged numerical results.

Refs #5757.

Coding agent: Codex
Codex version: codex-cli 0.144.4
Model: gpt-5.6-sol
Reasoning effort: xhigh