ci: shard bpf-tests by integration test binary

[not-matthias]

Summary

The bpf-tests job is the long pole of CI at ~13min, while every other job finishes in under 5min. Breakdown of the slow step on a recent run:

stage	time
build/deps	~1m15s
c_tests	3m38s (8 cases, serial)
cpp_tests	3m44s (7 cases, serial)
rust_tests	2m21s (3 cases, serial)
spawn_tests	1m03s (2 cases)

Each memtrack integration test binary runs its cases serially because the eBPF tracker can't safely overlap with itself in one process (--test-threads 1). That's a per-binary constraint though — different test binaries are fully independent.

This shards bpf-tests into four matrix jobs, one per integration test binary. Each shard builds only the binary it needs and runs its cases as before. Critical path drops from ~13min to ~5min (longest shard ≈ cpp_tests).

Test plan

• All four bpf-tests (...) matrix jobs pass
• check job correctly waits for the whole matrix and turns green

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

✅ 7 untouched benchmarks

---

_{Comparing ci/shard-bpf-tests (942f43e) with main (149a618)}