test: split proc stat limit coverage by architecture

[happysnaker]

Fixes #432.

This keeps the ProcStat limit coverage architecture-aware without relying on the shared 64-bit fixture values in testdata/fixtures.ttar.

What changed

• move TestProcStatLimits out of proc_stat_test.go
• add 32-bit and 64-bit test entry points with the same build-tag split pattern used by proc_maps*_test.go
• generate a tiny temporary /proc/<pid>/stat fixture per architecture, so the parsed CUTime / CSTime values match the native int width of that build target

Why this shape

The earlier math.MinInt / math.MaxInt constant-only approach still leaves a mismatch with the shipped fixture data, which encodes 64-bit extrema for those fields. On 32-bit architectures that means the test can compile but the parsed values are still not representable as native ints.

Using a temporary stat fixture keeps the test focused on the parser behavior we want to validate on each target architecture, while leaving the shared repository fixtures unchanged.

I did not run go test here; this patch is based on the issue discussion and static review only.