[DOCS][TIRX] Add in-kernel profiling (CudaProfiler) tutorial

[spectrometerHBH]

This adds an in-kernel profiling page to the TIRx native-basics CUDA section,
documenting the existing tvm.tirx.bench.CudaProfiler.

The page covers:

• a minimal load / compute / store example using start / end / finalize
  markers and a user-supplied uint64 buffer;
• decoding the record buffer on the host and exporting a Perfetto trace via
  export_to_perfetto_trace;
• the record/tag encoding and the device code each call lowers to (a
  %globaltimer read, a leader-only global store, and a block fence);
• usage notes: one leader per (block, group), buffer sizing, the 32-bit
  %globaltimer wrap, and the per-region cost.

The example is tested end-to-end on a CUDA GPU (B200, sm_100). It is wired into
the native_basics.rst toctree after "Compiling and inspecting". The
FlashAttention-4 timeline screenshot is served from tlc-pack/web-data
(images/tirx/tirx_cudaprofiler_fa4.png), matching the other tirx doc figures.

[gemini-code-assist]

Code Review

This pull request introduces documentation for in-kernel profiling using CudaProfiler in TVM's TIRX. The review feedback highlights several critical issues in the provided Python code examples, including the use of standard Enum instead of IntEnum (which causes TVM FFI compatibility issues), the use of the non-existent tvm.runtime.tensor API instead of tvm.nd.array, and a potential KeyError during trace decoding when handling unmatched end events.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Import IntEnum instead of Enum to support the integer-based enum definition for Ev.

Suggested change

	from enum import Enum
	from enum import IntEnum

[gemini-code-assist]

Using enum.Enum for Ev might cause TVM compilation or FFI errors because standard Enum members are not instances of int and cannot be automatically converted by TVM's FFI. Changing Ev to inherit from enum.IntEnum ensures that the enum members behave as integers and are compatible with TVM Script and FFI.

Suggested change

	class Ev(Enum):
	Load = 0
	Compute = 1
	Store = 2
	class Ev(IntEnum):
	Load = 0
	Compute = 1
	Store = 2

[gemini-code-assist]

In TVM, tvm.runtime.tensor is not a standard API and will raise an AttributeError. Use tvm.nd.array instead to allocate and initialize the NDArrays.

Suggested change

	inp = tvm.runtime.tensor(np.ones(N, "float32"), device=dev)
	out = tvm.runtime.tensor(np.zeros(N, "float32"), device=dev)
	prof = tvm.runtime.tensor(np.zeros(PROF_SIZE, "uint64"), device=dev)
	inp = tvm.nd.array(np.ones(N, "float32"), device=dev)
	out = tvm.nd.array(np.zeros(N, "float32"), device=dev)
	prof = tvm.nd.array(np.zeros(PROF_SIZE, "uint64"), device=dev)

[gemini-code-assist]

If the profiler buffer wraps or contains incomplete trace data, an end event might be processed without a matching start event in opens. Using opens.pop((block, event_idx), None) avoids a potential KeyError and safely ignores unmatched end events.

Suggested change

	elif event_type == 1:
	spans.setdefault(block, []).append((EV_NAMES[event_idx], ts - opens[(block, event_idx)]))
	elif event_type == 1:
	start_ts = opens.pop((block, event_idx), None)
	if start_ts is not None:
	spans.setdefault(block, []).append((EV_NAMES[event_idx], ts - start_ts))