fix(anthropic): cached token double-counting in OpenAI-to-Anthropic response transformation

[praxis-bot]

build_usage in apis/src/anthropic/to_openai/response.rs sets Anthropic
input_tokens to OpenAI prompt_tokens (which includes cached tokens),
then separately sets cache_read_input_tokens. In Anthropic's API contract,
input_tokens excludes cached tokens.

Downstream parse_anthropic in token_usage/providers.rs sums
input_tokens + cache_read_input_tokens, double-counting cached tokens.
Clients receiving this transformed response also see inflated input_tokens.

Impact: Incorrect token counts in billing dashboards and cost tracking
for Anthropic-translated responses.

Fix: Subtract cached_tokens from prompt_tokens when building
input_tokens:

let input_tokens = if let Some(cached) = cache_read {
    prompt_tokens.saturating_sub(cached)
} else {
    prompt_tokens
};

[mkoushni]

/assign

[leseb]

@mkoushni please review the bug report and leave a comment like this:

Reviewed by: @mkoushni 
Verdict: <Valid | Rejected>
Rejected: if applicable

Then I'll set the label "accepted" to this bug report. Thanks!

[mkoushni]

Reviewed by: @mkoushni
Verdict: Valid