ggml-webgpu: Support GPU profiling beyond the maximum query count

[yomaytk]

Overview

This PR fixes the bug described in the Additional Information section.

• Flush timestamp slots and reset the timestamp state when the number of used timestamp slots is nearly full.

I confirmed that GPU profiles can now be collected for Qwen3.5-35B-A3B-GGUF and several other models (Qwen3.5, Qwen3.6, Gemma 4, and Llama 3).

Additional Information

I noticed that unsloth/Qwen3.5-35B-A3B-GGUF overflowed the timestamp QuerySet when I tried to collect a GPU profile:

llama.cpp/ggml/src/ggml-webgpu/ggml-webgpu.cpp:571: GGML_ASSERT(ctx->profile_timestamp_query_count + 2 <= WEBGPU_MAX_PROFILE_QUERY_COUNT) failed

This suggests that we need logic to allow profile collection even when a model requires more than 4096 timestamp queries.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - I used AI to investigate WebGPU specification

[reeselevine]

thanks, this is a nice clean addition!