UCP: Enable ppln protos with cuda buffers by default

[brminich]

What

Enable using cuda staging buffers for pipeline protocols by default.

Why ?

Using ppln protocols with cuda staging buffers provides better performance for the following cases:

• sending data from cuda-managed memory to cuda memory
• sending data between different nodes where MNNVLK is available
• etc

How ?

Rename RNDV_FRAG_MEM_TYPE to RNDV_FRAG_MEM_TYPES and allow to specify a list of memory types. Then initialize RNDV_RECEIVE protocol variants (rtr_mtype and get_mtype) with these memory types. RNDV_SEND protocol has to use the mem type which corresponds to the rkey from received RTR, because:

• We use different fragment sizes for different mem_types, so rtr_mtype and put_mtype sizes will be different which breaks the protocol
• Currently there is no strong use-case to support ppln protocols with different mem_type fragments being used by peers