bindings/rust: bind the optional-args custom parameter (#2155)

## What?

Binds the optional-args custom parameter into the Rust C API layer and
`nixl-sys`.

- C API: `nixl_capi_opt_args_set_custom_param` / `_get_custom_param`,
with
  matching `stubs.cpp` forwarders
- nixl-sys: `OptArgs::set_custom_param` / `get_custom_param`
- a memory view test that sets `worker_id=0` and prepares a view with it

## Why?

`prepMemView` is the reason. `nixlAgent::prepMemView` copies
`extra_params->customParam` into the backend args in both overloads
(`nixl_agent.cpp:1921`, `:1994`), and the UCX backend parses
`worker_id=<n>`
out of it to choose which shared worker the view binds to
(`ucx_backend.cpp:1036-1060`). The device API test sets it the same way:

```cpp
extra_params.customParam = "worker_id=" + std::to_string(worker_id);
// test/gtest/device_api/single_write_test.cu:455
```

Without a binding a Rust caller cannot select the worker, so the memory
views
added in #2060 are reachable but not fully controllable. It is the only
field
of `nixl_opt_args_t` that `prepMemView` reads, and it was the one field
the
bindings did not expose.

## How?

The setter is byte-oriented (`const void *`, `size_t`), mirroring
`nixl_capi_opt_args_set_notif_msg` rather than the `const char *` of
`set_ip_addr`. The blob is not always text: the gpunetio backend stores
a raw
`cudaStream_t` in it (`gpunetio_backend.cpp:1141`), which a C-string API
would
truncate at the first zero byte.

## Testing

106 tests pass with `cargo test -- --test-threads=1` — 97 in `tests.rs`,
5 in
`mem_view.rs`, 4 in `test_sync_manager.rs` — on UCX 1.21 with a CUDA
device.
The new test runs by default; it needs a GPU but no device-capable RDMA
lane.


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Added support for setting and retrieving custom binary parameters
through optional arguments.
* Custom parameters preserve arbitrary byte content, including embedded
zero bytes.
* Added support for using custom parameters when preparing local GPU
memory views.

* **Bug Fixes**
* Improved handling of returned binary data, including empty values and
notification messages.

* **Tests**
* Added coverage for parameter round-tripping and CUDA-based memory view
preparation.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Shahaf Kiselnik <smeir@nvidia.com>