fix: preserve cached and reasoning tokens without double billing

[kevin-lozada-santos]

Cache-read and reasoning counts were dropped before reaching pricing. This inflated the issue's OpenAI-compatible high-cache fixture from 156,204 to 461,258 microdollars; native Anthropic instead lost the cache-read charge.

Normalize SDK responses and flat agent-step/dispatch counts into disjoint Usage buckets exactly once. input excludes cached reads and output excludes reasoning; flat ModelResponse/step counts remain inclusive totals. Populate both bundled adapters, account for native Anthropic's already-disjoint input, and preserve total prompt size in context-compaction trends. Document the contract and migration in Usage, ModelResponse, and the changelog. Pricing itself is unchanged.

Validation:

• 13 new offline tests cover Chat Completions, Responses, native Anthropic, both actual adapters with mocked SDK calls, flat responses/steps, zero/missing details, legacy responses, reasoning pricing, the reported high-cache fixture, and rising cached context. 11 failed before implementation and all 13 pass afterward.
• Full default python -m pytest -q: 244 passed, 43 skipped, 1 live test deselected; one dependency deprecation warning.
• ruff check src tests examples, ruff format --check src tests examples, and mypy src/tokenops pass (58 source files).
• git diff --check passes.

No live provider calls or original SWE-bench rerun. Cache-write premiums, streaming usage, additional adapters and context-compaction capabilities remain out of scope. The change to the compaction detector only retains its existing total-context semantics after separating cache reads.

Fixes #137. Follows the provider-aware approach discussed and approved in the issue.

[kevin-lozada-santos]

The implementation and all local default tests/lint/type checks are complete. GitHub is waiting for maintainer approval to run the CI workflow for this contribution. Please enable it when reviewing the PR; I can address any resulting failures.