[Make] Add USE FLASHINFER into doc and gen cmake config

[CharlieFRuan]

When using CUDA with compute capability >80, it is required to build runtime with USE_FLASHINFER and specifying FLASHINFER_CUDA_ARCHITECTURES and CMAKE_CUDA_ARCHITECTURES. Otherwise, users may run into issues like Cannot find PackedFunc flashinfer.single_prefill in either Relax VM kernel library, or in TVM runtime PackedFunc registry, or in global Relax functions of the VM executable.

This is related to issues #1728 and #1551

[tqchen]

cc @yzh119

[CharlieFRuan]

For more context: when compiling a model, the optimization flag flashinfer is true when the target is CUDA and the compute capability is above 80:

mlc-llm/python/mlc_chat/interface/compiler_flags.py

Lines 59 to 73 in 9ddffd6

	def _flashinfer(target) -> bool:
	from mlc_chat.support.auto_target import ( # pylint: disable=import-outside-toplevel
	detect_cuda_arch_list,
	)
	if not self.flashinfer:
	return False
	if target.kind.name != "cuda":
	return False
	arch_list = detect_cuda_arch_list(target)
	for arch in arch_list:
	if arch < 80:
	logger.warning("flashinfer is not supported on CUDA arch < 80")
	return False
	return True

Otherwise, we prune create_flashinfer_paged_kv_cache so that corresponding methods do not get looked up in runtime:

mlc-llm/python/mlc_chat/compiler_pass/prune_relax_func.py

Lines 23 to 28 in 9ddffd6

	for g_var, func in mod.functions_items():
	# Remove "create_flashinfer_paged_kv_cache" for unsupported target
	if g_var.name_hint == "create_flashinfer_paged_kv_cache" and not self.flashinfer:
	continue
	func_dict[g_var] = func
	ret_mod = IRModule(func_dict)

Therefore, if users use a "qualified" machine to compile the model, but the user does not compile TVM runtime with USE_FLASHINFER ON explicitly, then they would run into Cannot find PackedFunc issues.