serve: add /metrics endpoint for instance and cache size info

[tinder-maxwellelliott]

What

Adds GET /metrics to the serve query service so callers and monitoring can see a running instance's identity, liveness, and cache size usage without scraping logs.

Example:

curl 'http://localhost:8080/metrics'

{
  "version": "31.4.0",
  "uptimeSeconds": 3600,
  "ready": true,
  "gitEngine": "jgit",
  "trackDeps": false,
  "cache": {"directory": "/var/cache/bazel-diff", "entries": 128, "sizeBytes": 4823913, "sizeHuman": "4.6 MB"},
  "jvm": {"usedBytes": 123456789, "maxBytes": 2147483648}
}

How

• MetricsService assembles the snapshot from the instance's config, its readiness flag, live JVM state, and — when the backing store supports it — the on-disk cache footprint. Read on demand per request (metrics are polled at a low rate, so walking the cache dir each call is fine).
• MeasurableHashCacheStorage capability interface (mirrors the existing pluggable-storage design). LocalDiskHashCacheStorage reports entry count + total bytes by summing its *.json entries. A backend that can't cheaply report size (a future S3 store) simply doesn't implement it, and the cache size fields are null.
• Wired into BazelDiffServer as an always-on endpoint. Deliberately not gated on readiness — a scrape of an un-ready or lame-ducked instance still returns data, with the current state in the ready field.
• No new CLI flags.

Notes for reviewers

• Independent of the cache-pruning PR (serve: prune cached hashes to bound cache growth #417) — based off master. There is a small, intentional overlap in HashCacheStorage.kt: this PR adds a MeasurableHashCacheStorage capability while serve: prune cached hashes to bound cache growth #417 adds a PrunableHashCacheStorage one. They're separate interfaces/methods and will reconcile trivially when both land.
• sizeHuman uses base-1024 units to match the --cacheMaxSize scale from serve: prune cached hashes to bound cache growth #417.

Testing

• Unit: MetricsService snapshot assembly (injected clock for deterministic uptime; null cache for a non-measurable backend), and LocalDiskHashCacheStorage.stats().
• HTTP: /metrics returns 200 + JSON, 405 on non-GET, 404 when no provider is wired, and is served even when not ready.
• End-to-end: a ServeCommandTest case that stands up the real server via buildAndStartServer and curls /metrics over real HTTP, reading live on-disk cache stats parsed from JSON.
• ktfmt/buildifier clean; docs in readme_template.md + regenerated README.md.

Note: a full real-binary smoke of serve requires a real Bazel workspace (its DI eagerly runs bazel info at startup, unrelated to metrics), which is what E2ETest#testServeEndToEnd covers; the endpoint here is exercised end-to-end via the in-process ServeCommandTest over real HTTP.