feat(tracing): report cached input token usage

[evanlowe]

LLM token metrics currently expose input and output usage but omit cache hits, even when model response metadata includes cached input tokens

This adds gen_ai_token_type=cache_read to the existing gen_ai.client.token.usage metric and stops reporting cache reads as cache creation in traces

• Keep input and output totals unchanged: 1,000 input tokens with 600 cached tokens report input=1000 and cache_read=600
• Record an explicit zero when provided; omit cache-read samples when the normalized cache count is absent
• Cache reads are a subset of input usage, so total-consumption queries must continue selecting input and output rather than summing every token type

Validation: pre-commit passed and 57 related tests passed, including Volcengine and BytePlus endpoint labels, streaming and non-streaming recording, zero or missing cache counts, and trace attributes

Validation uses local response fixtures and an in-memory OpenTelemetry reader; no live cloud calls were made