[FEA] Allow RMM_DEFAULT_HOST_ALIGNMENT alignment in pool_memory_resource

[PointKernel]

Is your feature request related to a problem? Please describe.
Currently, the alignment check in pool_memory_resource ctor is hardcoded to check against CUDA_ALLOCATION_ALIGNMENT:

rmm/include/rmm/mr/device/pool_memory_resource.hpp

Lines 138 to 141 in 32cd537

	RMM_EXPECTS(rmm::is_aligned(initial_pool_size, rmm::CUDA_ALLOCATION_ALIGNMENT),
	"Error, Initial pool size required to be a multiple of 256 bytes");
	RMM_EXPECTS(rmm::is_aligned(maximum_pool_size.value_or(0), rmm::CUDA_ALLOCATION_ALIGNMENT),
	"Error, Maximum pool size required to be a multiple of 256 bytes");

This causes problems for host memory pools.

Additional context
Related issue: rapidsai/cudf#16024

[harrism]

I discussed this internally at NVIDIA, and the current recommendation for anything device-accessible is to leave the required alignment at rmm::CUDA_ALLOCATION_ALIGNMENT, and not relax it to DEFAULT_HOST_ALIGNMENT. I'll leave this issue open for discussion for now.