Pass the memory handler and filter config to exclude cuda transport

[tvegas1]

Due to parallel cuda kernel running, UCX plugin must not try to use cuda primitives to manipulate passed cuda memory. To do so we:

1. Disable cuda transport alias altogether
2. Don't try to resolve the cuda memory chunk to the whole cuda memory area
3. Force RNDV to avoid the need for explicit memory copy (eager, rx path)
4. Pass memory handle to allow memory details to be looked-up instead

[tvegas1]

pls look @brminich, @yosefe

[brminich]

lgtm, but can we make it configurable just in case? I.e. have a single NCCL var disabling or enabling cuda support in UCX plugin

[tvegas1]

lgtm, but can we make it configurable just in case? I.e. have a single NCCL var disabling or enabling cuda support in UCX plugin

as discussed, added complete bypass config possibility to prevent blocking any further testing down the line.