fix(deepseek): extract prompt_cache_hit_tokens and reasoning_tokens from usage

[Lintume]

Description

DeepSeek\Handlers\Text::addStep() and Stream::extractUsage() hardcode Usage to only prompt_tokens and completion_tokens from the API response, silently dropping two DeepSeek-specific fields:

• usage.prompt_cache_hit_tokens — cached input portion of the prompt. DeepSeek offers a 98% discount on cache hits (their headline feature) and exposes this as a separate counter.
• usage.completion_tokens_details.reasoning_tokens — internal thinking tokens emitted by reasoning models (`deepseek-reasoner`, `deepseek-v4-flash` thinking mode).

Impact

• Apps that compute cost from Prism's `Usage` charge the full `prompt_tokens` at fresh rate — overstating real spend ~3-5x once the prompt cache warms up.
• No signal to derive a `cacheHitRatio` for monitoring prompt-prefix stability.
• Reasoning-mode token consumption is invisible to observability tooling (Langfuse, custom dashboards, etc.).

Changes

• `src/Providers/DeepSeek/Handlers/Text.php`::addStep() — read both fields from `usage`, subtract cache hit from `prompt_tokens` to derive the fresh-prompt count, populate `Usage` with `cacheReadInputTokens` and `thoughtTokens`. Mirrors what the Gemini and OpenAI handlers already do for their analogous fields.
• `src/Providers/DeepSeek/Handlers/Stream.php`::extractUsage() — same fix in the streaming path.
• `tests/Providers/DeepSeek/TextTest.php` — multi-step tools test updated to assert the new semantics (fresh-only `promptTokens` aggregated across steps, plus the previously-invisible `cacheReadInputTokens`).

All 28 DeepSeek tests pass. Pint clean.

Testing

Verified against `deepseek-v4-flash` direct API in production:

• Cold cache request → `cacheReadInputTokens=0`, `promptTokens=`
• Warm cache (same prefix within ~hour TTL) → `cacheReadInputTokens=`, `promptTokens=`
• Reasoning model → `thoughtTokens` populated and matches the DeepSeek platform dashboard

References

• DeepSeek pricing — cache hit / miss tokens
• DeepSeek context caching docs

Non-overlapping with #1020 (which fixes `reasoning_content` round-trip) — both fixes are complementary; this one is purely about exposing usage detail that's already in the response.