Add LLImpl pack/unpack over LLPacket geometry (#3554)

[function47]

Summary:

Second llx building block: LLImpl<P>, the ThreadGroup-driven encode/decode of
a byte range into/out of an LLPacket-formatted staging region.

• pack(group, staging, src, nbytes, flagVal): encode the payload into
  consecutive packets, stamping each packet's trailing flag with flagVal.
• unpack(group, dst, staging, nbytes, flagVal): spin until each owned
  packet's flag == flagVal, then decode the payload out.
• Flag I/O (store_flag / load_flag / is_flag_set): the flag sits entirely
  inside P::kFlagLane's 16 B slot, so it is read/written as ONE wide volatile
  transfer instead of a per-word scalar loop. flagVal is replicated across the
  full flag width, so a torn transfer that carries the new value in only part of
  the flag fails is_flag_set — unpack never accepts half-arrived data.

Parallelism (v1): one thread owns one whole packet, grid-strided across the
group. pack writes each packet's trailing flag last; on the wire the whole
staging region is fenced and RDMA-put as one unit, so the flag is never observed
before its data. unpack spins only on the packets a thread owns (no
group-wide readiness barrier), then decodes.

Additive only: new header comms/prims/core/LLImpl.cuh behind a new
//comms/prims:ll_impl target, plus //comms/prims/tests:ll_impl_test and its
:ll_impl_test_kernels companion. No existing code path is modified.

Reviewed By: snarayankh

Differential Revision: D110940544

[meta-codesync]

@function47 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D110940544.

[meta-codesync]

This pull request has been merged in a66c6f6.