feat: Add a Python Callout implementation for the Prompt Amendment / Insertion

[victorwoli]

Prompt-Amender: ext_proc callout for centralized LLM prompt governance

Adds prompt_amender, an ext_proc callout that matches requests by caller identity/host/path against a hot-reloadable YAML ruleset and mutates systemInstruction in flight — injecting safety guardrails and context without client code changes or redeploys.

Highlights

• Header-phase glob matching (spiffe://``/principalSet:// normalized); mode_override skips body buffering for unmatched traffic.

• Body-phase mutation (prepend/append/replace/template, sandboxed Jinja2, compiled once); accepts both systemInstruction and system_instruction, inserting one if missing.

• Correctly handles Envoy's streamed body delivery — non-final chunks withheld via clear_body rather than naively accumulated, so multi-chunk requests aren't corrupted; failures reconstruct the original body (fail-open) or reject (fail-closed).

• Hot-reload from GCS/git with atomic swap and validation-reject.

• Structured JSON logs (no raw prompt text) and real OTEL metrics/tracing.

Included: callout code, Terraform (with strip_client_spiffe_id defaulting true — identity headers only trustworthy behind a real gateway), Dockerfile, README, and 38 unit tests (no gRPC/network deps).

Follow-up: callouts/python/requirements-test.txt needs Jinja2==3.1.4 added (see requirements-test-addition.txt) — a shared file outside this PR's diff.