Add critical-path NVTX annotations

[9prady9]

Summary

• add coarse NVTX ranges for Sirius-relevant GPU, host, and disk conversions
• expose reservation, batch-lock, and reader-event waits in the libcucascade domain
• keep uncontended reservation and lock paths quiet, with numeric payloads for requested bytes and pending reader events

Motivation

Sirius already annotates queries, pipelines, tasks, and operators, but cuCascade's memory-pressure and tier-transition work is mostly opaque beneath those ranges. This adds a restrained first layer of library-side visibility without annotating allocators, fast try-lock paths, individual buffers, or worker-loop lifetimes.

Conversion labels spell out both endpoints and group lexically under convert: in the profiler UI:

• convert:gpu_to_host / convert:host_to_gpu / convert:gpu_to_gpu
• convert:host_to_disk / convert:disk_to_host
• convert:gpu_to_disk / convert:disk_to_gpu

Existing host-to-GPU phase ranges nest under convert:host_to_gpu:*.

Validation

• pixi run -e cuda-13-nightly build
• focused GPU↔HOST and GPU/HOST↔DISK conversion benchmarks
• Nsight Systems actual-library smoke confirmed libcucascade:wait:reservation and libcucascade:wait:batch_lock with expected ~20 ms contention lifetimes
• inspect the conversion ranges in a representative Sirius Nsight capture

Scope

This intentionally leaves generic converter wrappers, packed legacy converters, detailed phase breakdowns for new directions, and I/O/reactor annotations for a later profiling-driven pass.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[9prady9]

@felipeblazing @aminaramoon pls provide inputs on added annotations and suggest any missing locations. ty

[mbrobbel]

/ok to test c9adf6b