bench: resize the microbenchmarks that flip on unrelated pull requests - #9951

…9994)

## Summary

Mimalloc's deferred page collection and arena purging can land inside a
single-shot benchmark and get reported as a Vortex regression. In [this
CI
comparison](https://app.codspeed.io/vortex-data/vortex/runs/compare/6ab29cf078100e55aedd4885..6ab2a14f78100e55aedd4921),
`_mi_theap_collect_retired` and its callees account for an additional
22.30 µs in the dictionary case and 38.64 µs in the FSL case, closely
matching their reported gaps.

## Changes

Enable `exclude-allocations` for the array dictionary and selection
simulation shards, which contain `dict_compress`, `take_fsl`, and
`take_chunked`. This keeps mimalloc in use while excluding allocator
time from these scores. Allocation-focused benchmarks in the core shard
continue to measure allocator cost.

<details>
<summary>Previous attempts</summary>

- #8742 and #8861 addressed earlier two-level CodSpeed results by adding
mimalloc. #8807 tracked the cases that remained flaky.
- #9951 resized noisy cases, and #9953 extended mimalloc to the
remaining benchmark binaries. Deferred cleanup still varies within
mimalloc.
- #9174 included a commit titled “Try and exclude codspeed allocations,”
but its diff only upgraded the action. Its report used runner 5.0.1,
where [allocation exclusion was already disabled by
default](CodSpeedHQ/codspeed#489). No explicit
opt-in was added. This PR sets it for the affected shards.

</details>

Signed-off-by: Connor Tsui <[email protected]>

…9980)

## Summary

The decimal kernel benchmarks landed with a 1,024 row case. For `i128`
that runs in 0.4 to 1.2 µs
on the walltime legs, below what those machines resolve, so it reports
mostly jitter: it moved by
10 to 19% on ten of the last thirteen CodSpeed reports, every one of
them on a pull request that
touched no decimal code. No 8,192 row case and no `i256` case has moved
in those reports.

## Changes

Derive each kernel case's row count from a working set of 256 KiB and 1
MiB of input plus output,
rather than fixing it at 1,024 and 8,192 rows:

| Decimal width | Small case | Large case |
| --- | ---: | ---: |
| `i128` | 8,192 rows | 32,768 rows |
| `i256` | 4,096 rows | 16,384 rows |

Both widths now do comparable work, every case takes several
microseconds instead of fractions of
one, and the largest still fits the 1 MiB L2 cache of the Graviton leg,
so these stay measurements
of kernel code rather than of memory bandwidth.

The array benchmarks in the same files are `#[cfg(not(codspeed))]` and
never run in CI, so they
keep their sizes for quick local runs. The benchmarking guide gains the
measured walltime floor
next to the simulation floor documented in #9951.

Expect a one-time batch of CodSpeed rows for these eight cases when this
merges, since the inputs
change size.

Checked with `cargo clippy -p vortex-decimal-byte-parts --all-targets
--all-features -- -D warnings`
and `cargo fmt`. No benchmarks were run locally: this session is a
shared cloud VM, so its walltime
numbers would not mean anything.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01B6ApyxEXpjiasZbQptsypn


---
_Generated by [Claude Code](https://claude.ai/code)_

---------

Signed-off-by: Connor Tsui <[email protected]>