ciflow/trunk/182407: Update on "Support memory snapshot for CPU pinned memory"

This adds memory profiling support for the CPU pinned memory (host)

allocator, so that torch.cuda.memory._snapshot() can include pinned

memory allocations alongside CUDA device memory.

To enable, pass record_host=True to _record_memory_history().

Host data appears in the host_segments and host_traces keys of

the snapshot dict. Disabling with _record_memory_history(None)

turns off both CUDA and host recording. The memory_viz visualizer

does not yet support host memory data.

Implementation: the CachingHostAllocatorImpl gains the same

profiling infrastructure as the CUDA DeviceCachingAllocator:

context gathering, trace recording via a RingBuffer, and

getSegments()/getTraces() APIs. The RingBuffer class is

extracted from CUDACachingAllocator.cpp into c10/core/RingBuffer.h

for reuse. Timestamp conversion is done inside getTraces() rather

than at snapshot collection time.

Authored by Claude.

Signed-off-by: Edward Z. Yang <ezyangmeta.com>

[ghstack-poisoned]