Trace bounded NVTX range boundaries

[itdevwu]

Goal

Use NVTX ranges as explicit framework and application boundaries without expanding into CUPTI metric replay.

Acceptance criteria

• Public selectors represent NVTX range start and end with bounded name matching and explicit nesting or range-ID semantics.
• validate reports whether the running process can be observed after online CUPTI injection; unsupported NVTX initialization states fail explicitly.
• Collection enables only required NVTX activity or callback sources and copies only bounded range identifiers and names.
• Extended payloads, pointer-referenced arguments, and environments are not collected by default.
• CUDA 12/13 and PyTorch fixtures cover thread ranges, cross-thread ranges where supported, nesting, drops, and cleanup.
• Documentation distinguishes NVTX timeline ranges from CUPTI Range Profiling metrics and replay.