fix(env): cache MORI_ENABLE_SDMA / MORI_DISABLE_P2P at Context construction

[jhchouuu]

Problem

MORI_ENABLE_SDMA and MORI_DISABLE_P2P were read independently from multiple call sites:

Location	When read	What it controls
Context::InitializePossibleTransports	once at shmem init	transport selection (P2P / SDMA / RDMA)
SymmMemManager::Malloc	every allocation	hipMalloc vs hipExtMallocWithFlags(hipDeviceMallocUncached)
EpDispatchCombineHandle ctor	every handle construct	config.enableSdma flag

If MORI_ENABLE_SDMA was set after Context init (e.g. a session-scoped pytest fixture spawns workers, workers init shmem, then the test function does os.environ["MORI_ENABLE_SDMA"] = "1"), the transport layer still believed P2P but per-allocation paths flipped to uncached buffers — producing a cache/IPC inconsistency that silently hung async_ll.

This was the latent reason tests/python/ops/test_dispatch_combine_async_ll.py had to be invoked with the env on the pytest command line in .github/workflows/ci.yml (PR #313); setting it inside the test function had no effect on transport selection but did flip allocation cacheability — the worst combination.

Fix

Snapshot both env vars exactly once in Context's constructor and expose them through Context::IsSdmaEnabled() / Context::IsP2PDisabled(). All later readers consult the cached value:

• SymmMemManager::Malloc → context.IsSdmaEnabled() (already had a Context& reference)
• EpDispatchCombineHandle → new mori::shmem::ShmemSdmaEnabled() accessor that delegates to Context::IsSdmaEnabled()

After this PR: either the env is set before shmem_init and takes effect end-to-end, or it isn't and is silently ignored — there is no half-on state.

Files

• include/mori/application/context/context.hpp — add bool sdmaEnabled / bool p2pDisabled members + getters
• src/application/context/context.cpp — initialize them in ctor; switch internal sites to member accessors; remove the file-local free functions
• src/application/memory/symmetric_memory.cpp — SymmMemManager::Malloc consults context.IsSdmaEnabled()
• include/mori/shmem/shmem_api.hpp + src/shmem/runtime.cpp — add ShmemSdmaEnabled() accessor (mirrors existing ShmemNumQpPerPe())
• src/ops/dispatch_combine/dispatch_combine.cpp — EpDispatchCombineHandle ctor uses ShmemSdmaEnabled()

Test plan

• CI green on MI300X-BNXT and MI355X-AINIC (this PR's box is MI308, SDMA queue init flaky on this hardware so local async_ll over SDMA isn't a reliable signal)
• intranode and internode_v1 jobs unchanged (no SDMA path involved → no behavior change)
• async_ll job (which sets MORI_ENABLE_SDMA=1 on the pytest cmdline) keeps passing — confirms the cached path is wired correctly