fix(anthropic): capture input + cache tokens on the streaming path (#433)

[initializ-mk]

Fixes #433.

Problem

readAnthropicStream handled only message_delta (which carries output_tokens), so a streamed Anthropic call dropped all input tokens — and, under prompt caching, the cache_read/cache_creation counts too. The streaming sibling of the non-streaming fix in #431/#432.

Anthropic reports input_tokens (+ cache_read_input_tokens / cache_creation_input_tokens) on the message_start event and accumulates output_tokens onto message_delta — so a parser that ignores message_start sees zero input.

Fix

• Parse message_start usage (input + cache read/creation) into locals.
• Emit one complete UsageInfo on the terminal message_delta (input + cache read + creation + output + TotalTokens), so the streamed usage matches the non-streaming path.

Why one terminal emission (not emit-at-start + emit-at-delta): consumers merge StreamDelta.Usage differently — the existing aggregator (responses.go) overwrites (result.Usage = *delta.Usage), while #433's sketch assumed summing. A single authoritative Usage is correct under both: overwrite ends with the complete value; summing sees it once (no double-count). Emitting input at start and output at delta would lose input under the overwrite consumer.

Blast radius

Low/forward-looking: today the only production ChatStream consumer (forge-cli/cmd/ui.go) reads Content only and ignores usage, so nothing consumed streaming usage before. This makes it correct for when a consumer (or a future streaming AfterLLMCall/accumulator) does.

Tests

• Cache-heavy stream recovers input=12, cache_read=4000, cache_creation=200, output=25, total=4237, TotalInputTokens()=4212.
• Non-cached stream populates input+output with zero cache fields.

golangci-lint clean; full forge-core/llm suite passes. No doc change — the fix makes streaming match the usage behavior the docs already describe for llm_call.