[Abhishek-Varma]

Summary

Fixes alignment, granularity, and rollback edge cases bugs in VMemChunkedAllocator (refer #600)

Fixes

1. Granularity reporting - distinguish minimum vs. recommended allocation
   granularity.

   • Added BaseDriver.get_recommended_granularity() (a concrete default that
     returns the minimum) and query both minimum/recommended from the AMD
     local and fabric drivers.
   • The allocator now backs memory in units of the recommended granule
     while reporting the true minimum via get_stats()
     (recommended_granularity / min_granularity).
   • AMD behavior is byte-identical: the old get_minimum_granularity()
     already returned the recommended value.

2. Alignment - honor a requested alignment larger than the granule.

   • allocate() computes an effective alignment (≥ min_alignment, rounded
     to a power of two) and aligns the heap offset to it, keeping allocations
     symmetric across ranks.
   • Free lists are keyed by (size, alignment) - a Bucket - so a block is
     only reused for a request with matching alignment.

3. Use-after-close - add _check_open() guards to allocate(),
   import_external_tensor(), import_peer_chunk(), and
   get_allocation_chunks_since(). Operations after close() now raise
   RuntimeError instead of silently corrupting allocator state.

4. Failed-allocation rollback - build the tensor before committing the
   bump pointer and bookkeeping. If tensor construction raises, a popped
   free-list block is returned to its bucket and the bump pointer is left
   untouched - no orphaned reservation, so a rank can always re-converge to
   the symmetric layout.

5. Non-power-of-two heap sizes - drop the power-of-two chunk_size
   requirement. VA reservation is rounded up to a whole number of chunks
   arithmetically (math.ceil) instead of via a bitmask, so arbitrary
   granularity-aligned chunk sizes work for any heap size.

Testing

• Added reproduce-first regression tests for all five fixes:
  • Use-after-close coverage for all four guarded entry points (not just
    allocate()).
  • A free-list-path rollback test (in addition to the bump-pointer path).
  • A same-offset assertion proving the alignment fix actually reuses a freed
    block, rather than coincidentally landing on an aligned bump.
  • Two GPU-free driver-level unit tests pinning that
    get_minimum_granularity() / get_recommended_granularity() route
    distinct HIP flags.
• Full allocator suite: 37/37 passing on 2× MI300A (1- and 2-rank).
• Driver unit tests (GPU-free): 40/40 passing.