Add cgroup-cpuset-aware worker pinning to dbtest (--cgroup-aware)

[yans3meta]

Summary:
Silo's dbtest derives each worker's NUMA node from its logical core id
(effectively always node 0) and pins via numa_run_on_node(), so under a cgroup
v2 cpuset that excludes node 0 it aborts. This adds an opt-in --cgroup-aware
runtime flag and drops the explicit NUMA memory hint so this benchpress package
can run inside an off-node cpuset. Default behavior (no flag) is unchanged from
stock Silo.

The C++ change ships as a build-time patch
(packages/silo/patches/silo-cgroup-aware-no-numa.patch) applied during install
against the pinned upstream commit:

• allocator.cc: remove numa_hint_memory_placement()
  (mbind/numa_interleave_memory); the DB pool is placed by first-touch.
• core.{cc,h}: read the process CPU-affinity mask once via sched_getaffinity
  (reflects cgroup cpuset / numactl / taskset); add phys_cpu(),
  num_allowed_cpus(), and a cgroup_aware() runtime toggle.
• rcu.cc: when cgroup-aware, sched_setaffinity each worker to exactly its
  granted physical CPU (memory follows by first-touch); otherwise stock
  numa_run_on_node().
• bench.{cc,h}, dbtest.cc, ycsb.cc: plumb the --cgroup-aware flag; size loaders
  and workers off num_allowed_cpus() when set; require num-threads <= allowed
  CPUs under strict pinning; log the worker pin map.

Package wiring:

• install_silo.sh: git apply --check then git apply the patch after submodule init.
• README.md: document the no-NUMA-mempolicy difference and the --cgroup-aware flag.

Differential Revision: D109870007

[meta-codesync]

@yans3meta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D109870007.