[vm] Pin par_exec rayon workers to physical cores

[wqfish]

When the block executor builds its parallel execution pool and the
process has at least as many physical cores available as worker threads,
pin each worker 1:1 to a distinct physical core. This prevents two
workers from landing on HT siblings of the same core, which degrades
throughput on CPU-bound Block-STM workloads.

Approach

• Detect the physical cores the process may run on by grouping the CPUs
  from sched_getaffinity (via core_affinity::get_core_ids()) by
  their /sys/.../topology/thread_siblings_list, keeping one
  representative per group. Linux only; other platforms return None.
• Pin via a rayon start_handler that calls
  core_affinity::set_for_current using the worker's thread index.
• Skip pinning (fall back to the OS scheduler) when topology cannot be
  detected (non-Linux, sysfs unreadable, any read fails), the process
  has fewer physical cores than num_threads, or num_threads == 0.
• Per-thread pin failures are logged at warn! and the thread continues
  unpinned; the pool still comes up.

Robustness

• core_affinity::get_core_ids respects the process's affinity mask, so
  this composes correctly with cgroups, containers, and taskset.
• Unit tests cover the sibling-grouping logic (HT on/off, restricted
  affinity set, read-failure path).

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com