feat: add /compact endpoint and llm-d compaction service example

[OsherElhadad]

What

Makes context-guru usable as a standalone compaction service for
llm-d-router's
request-inline-compaction step (PR #1). That step POSTs the raw inference
request body to a compaction server and, on 200 + non-empty JSON, swaps that
body in. Contract: request body in → smaller body of the same shape out.

The proxy was forward-only (routes proxy to OpenAI/Anthropic and return a
completion), so this adds a compact-and-return endpoint plus the pieces needed
to run it cleanly without the store.

Changes

• POST /compact (proxy/): runs the pipeline over the request messages
  and returns the rewritten body without forwarding upstream. OpenAI by default,
  ?provider=anthropic; per-request overrides via ?preset= and
  x-context-guru-pipeline. Fail-open (bad input → original body, 200).
• Store enable/disable flag: store.enabled in YAML, plus --store /
  STORE. Adds a Nop store and Store.Persists(), so a full marker_mode
  automatically degrades to an irreversible drop when the store can't persist —
  instead of leaving an unresolvable <<cg:HASH>> marker (applied in
  mark/summarize/cmdfilter).
• toon component (components/reformat/): deterministic re-encoding of
  uniform JSON object-arrays into tabular TOON. Lossless repack; tests included.
• Config-driven model creds: the model: block on extract/summarize now
  takes provider/base_url/api_key/model/auth; empty api_key falls
  back to env keys (incl. ANTHROPIC_AUTH_TOKEN for bearer gateways).
• examples/llm-d-service/: three configs (toon, extract-code, summarize;
  store off, marker_mode: off), build.sh, a dependency-free Go client
  mirroring the router's call, a self-contained README, and a captured
  before/after sample run.

Testing

• CGO_ENABLED=1 go build ./..., go vet ./..., go test ./... — all pass;
  gofmt clean. New tests: toon encoder + the store-disabled marker degrade.
• End-to-end: ran the service for each config against a shared context.
  Deterministic toon reduced a sample request ~46%. The two LLM configs ran a
  real claude-haiku-4-5 call through a bearer gateway (extract filtered to
  the relevant lines deletion-only; summarize collapsed the transcript ~81%),
  all with the store disabled and no markers in the output.

Note

The commit skips the repo's dnephin go-vet pre-commit hook (SKIP=go-vet): it
passes individual filenames to go vet, which errors on a multi-package commit
("no Go files in ..."). go vet ./... is clean; all other hooks (gofmt,
go-mod-tidy, gitlint, DCO) ran.