perf: cache compiled CRD regex + skip hex-preview Vec allocation

[indyjonesnl]

Changes

Two unrelated perf wins, kept together because both are small and obviously safe.

`perf(schema-validation): cache compiled regex patterns`

`SchemaValidator::validate_string` recompiled `regex::Regex::new(pattern)` on every CRD field that declared a `pattern`. On a busy admission webhook this is a measurable hot path — regex compile dominates the validation cost.

Cache compiled patterns in a process-wide `LazyLock<RwLock<HashMap<String, Regex>>>`. Read-lock for the common hit path; write-lock only on miss. Bad regexes still return `InvalidResource` errors and are not cached.

Includes a microbenchmark (`#[ignore]`, run with `--release --ignored`) that compares the cached path against `Regex::new` per call:

```
regex_cache_speedup: n=5000 cached=3.59ms uncached=259.39ms speedup=72.3x
```

`perf(middleware): build hex body preview directly into String`

Protobuf debug-logging path used `.map(|b| format!("{:02x}", b)).collect::<Vec<_>>().join(" ")`, allocating an intermediate `Vec` of 80 entries on every protobuf request. Write directly into a pre-allocated `String` via `write!` — same output, one allocation instead of 81.

Verification

`cargo build --workspace --locked` clean. Two new tests on the regex cache (correctness + the ignored bench) pass. No new test failures vs baseline.

Depends on #32 for green CI.