feat(vpto): add sdma_gm_gm async DMA

[kerwinair]

Introduce the async communication micro-instructions on the VPTO path:

• !pto.dma_session<engine>: an opaque, engine-parameterised session type (sdma/urma/rdma) declared in ODS with CAPI and Python bindings.
• pto.session_init %template_gm -> !pto.dma_session<sdma>: materialises the device-side session from a host-built template in GM. The physical layout is pinned by include/PTO/Support/AsyncSessionABI.h (10 fields, width table); per-endpoint offsets are plain pointer arithmetic on the caller side rather than a session field.
• pto.sdma_gm_gm %dst, %src, %session {block_bytes, channel_idx, soft_put} -> !pto.ptr<i64, gm>: posts a GM->GM transfer and returns the channel record address, whose live head/tail words expose completion (*head == *tail). The A5 soft_put path writes synchronously and returns a null record. A2/A3 and A5 SQE layouts are both emitted.
• vpto-expand-wrapper-ops: an ordered walk lowers the opaque session to a !pto.struct plus ld_dev/st_dev sequences, then substitutes it into downstream users; LLVM emitter patterns for ld_dev/st_dev and byte-granular UB->GM config are ported onto the shared emitter helpers.

PTODSL exposes pto.dma_session_type(engine), pto.session_init and pto.sdma_gm_gm; @pto.jit can link host C++ into the kernel library so the host-side session template builder (test/comm/AsyncWorkspaceShim.cpp) is usable from Python tests.

Docs: docs/isa/micro-isa/19-async-comm.md, PTODSL user guide sections 4/7.

Tests: lit coverage for roundtrip, expansion, and A2/A3/A5 SQE emission (split per CHECK-LABEL so function-pass dump order cannot flake); PTODSL unit tests and docs-as-test fragments; ST cases under test/vpto/cases/async-comm/ (soft_put chunk tail, endpoint offset, multicore) and an engine-path spike under test/comm/. All verified on A5.