feat(vanity-gateway): expose dropped shadow request metric

[FamousDirector]

Problem

The vanity gateway records dropped shadow dispatches in trace attributes and a debug log, but it does not expose a Prometheus metric. Operators cannot quantify shadow drops or compare delivered shadow traffic with drops for each configured shadow model using metrics alone.

Proposed change

• Add a monotonic counter for dropped shadow dispatch attempts.
• Label the counter by openai_model_name and bounded drop reason: concurrency_limit, body_read_error, or body_rewrite_error.
• Pre-initialize every configured shadow-model and reason combination to zero so its series exists before the first drop.
• Add focused tests for counter initialization and per-model increments.

Acceptance criteria

• Prometheus exposes nvcf_ai_api_gateway_shadow_requests_dropped_total.
• Drops can be queried by openai_model_name and reason.
• Each recorded drop increments exactly one model and reason series.
• Multi-target failures increment the corresponding target-model series.
• Existing shadow dispatch and tracing behavior remains unchanged.
• Focused vanity-gateway tests pass.