[Relay] Fixed VNNI batch matmul op strategy for meta schedule compatibility

[masahi]

Currently, meta schedule task extraction for VNNI batch_matmul fails because

• When relay.backend.use_meta_schedule is True, is_auto_scheduler_enabled() is also True
• When is_auto_scheduler_enabled() is True, this branch

  tvm/python/tvm/relay/op/strategy/x86.py

  Lines 561 to 569 in 0009a30

  	if is_dynamic(out_type) or is_auto_scheduler_enabled():
  	strategy.add_implementation(
  	wrap_compute_batch_matmul(
  	topi.nn.batch_matmul, need_auto_scheduler_layout=True, need_out_dtype=True
  	),
  	wrap_topi_schedule(topi.generic.nn.schedule_batch_matmul),
  	name="batch_matmul.generic",
  	plevel=10,
  	)

  in op strategy is always taken, while I want to choose the x86-specific implementation of batch_matmul, topi.x86.batch_matmul in the else branch.

This PR fixes this problem by doing VNNI dispatch at the op strategy level before we enter is_auto_scheduler_enabled() == True branch, rather than inside topi.x86.batch_matmul.

cc @junrushao1994 This is my workaround for the issue we discussed last week.

[csullivan]

LGTM