Phase 2: Productionize linear_logp backend support and telemetry

[inaniloquentee]

Background

Split from #6 Phase 2: Productionize linear_logp.

This issue covers the production-readiness surface for the existing linear_logp integration: keep the hidden-state-to-selected-logprob path working, make supported configurations explicit, and report backend decisions with useful telemetry.

Scope

• Keep the existing hidden-state -> selected-logprob integration.
• Preserve the saved backward state required by the full-gradient path.
• Document the linear_logp support matrix:
  • dtype
  • hardware/backend
  • TP support
  • CP behavior
  • entropy support / fallback
  • full-gradient support
  • backend implementation name
• Emit structured runtime metadata for linear_logp decisions:
  • requested backend
  • actual backend
  • contract/backend ID when available
  • fallback status
  • fallback reason
  • operator-level timing
  • operator-level memory information when available

Acceptance Criteria

• Supported linear_logp configurations are discoverable from docs or runtime logs.
• Operator-level timing and memory claims remain separate from full-step performance claims.
• Existing native behavior is unchanged when the RL-Kernel fast path is disabled.
• The full-gradient path remains covered by tests or an explicit validation path.