ci: integrate callgrind benchmarks using iai-callgrind

[Zul-Qarnain]

This PR adds instruction-level benchmarks using iai-callgrind and integrates them into CI to catch regressions deterministically:

• Adds benches/callgrind.rs testing core vector operations (push, pop, insert/remove, slice conversions, drain, retain_mut, into_iter) across both inline and spilled heap storage.
• Adds a callgrind job to .github/workflows/checks.yml running the benchmarks on both default features and --all-features.

[alejandro-vaz]

CI shows nothing

callgrind::push_group::bench_smallvec_push_inline
  Instructions:                           0|N/A                  (*********)
  L1 Hits:                                0|N/A                  (*********)
  L2 Hits:                                0|N/A                  (*********)
  RAM Hits:                               0|N/A                  (*********)
  Total read+write:                       0|N/A                  (*********)
  Estimated Cycles:                       0|N/A                  (*********)

[alejandro-vaz]

it doesn't seem to be working properly

[Zul-Qarnain]

Fixed in 5cd5e5b.

The 0|N/A output was because cargo bench built the callgrind target with [profile.bench]'s debug = false + strip = true, so Callgrind had no function symbols to attribute events to and collected 0 for every benchmark.

I added a dedicated [profile.callgrind] (inherits bench, but debug = 1, strip = false, lto = false) and pointed both CI steps at it via --profile callgrind.

Verified on the same runner (fork Actions run): instruction counts are now non-zero, e.g. push_inline=361, push_spilled=2952, vec_push=120, pop_inline=484.

Note: the PR's own workflow is still sitting at the "Approve and run workflows" gate for fork PRs — could someone approve it so the callgrind check runs here too?

[alejandro-vaz]

does LTO have to be disabled in order for it to work??

[Zul-Qarnain]

No. I tested lto = "fat", "thin", and false (plus codegen-units 1 vs 16) across all 19 benchmarks — every config gives correct, non-zero counts within ~1% of each other (e.g. push_inline = 361 in all cases).

The actual requirement is just strip = false so Callgrind keeps the symbol table; neither debug nor lto affects whether events are collected. I've dropped lto = false in 701a661 so the callgrind run stays representative of the normal bench codegen.

[alejandro-vaz]

thanks for contributing