hexagon: add HTP backend support for GGML_OP_CLAMP

[cqderek]

CLAMP was falling through to default:break in device_supports_op(), causing the scheduler to dispatch it to CPU on every call. On MoE models (e.g. Qwen3-30B-A3B) this triggers 224 CPU<->HTP FastRPC round-trips per forward pass due to per-layer expert routing weight normalization, adding significant fixed latency in the decode path.

Fix: wire GGML_OP_CLAMP through the existing op_unary framework, reusing the already-present hvx_clamp_scalar_f32() HVX kernel. No new HVX vector code is introduced; all changes are plumbing only.

Benchmark (Q4_0, SM8850 HTP, NDEV=1, -t 4, tg50 x5):
Before: 18.61 +/- 0.17 tok/s (224 CPU splits/graph)
After: 20.24 +/- 0.28 tok/s (0 CPU splits for CLAMP)
Gain: +10.4%"

Overview

GGML_OP_CLAMP was falling through to default: break in device_supports_op(), causing the scheduler to dispatch it to CPU on every call. On MoE models (e.g. Qwen3-30B-A3B) this triggers 224 CPU<->HTP FastRPC round-trips per forward pass due to per-layer expert routing weight normalization, adding significant fixed latency in the decode path.

Fix: wire GGML_OP_CLAMP through the existing op_unary framework, reusing the already-present hvx_clamp_scalar_f32() HVX kernel. No new HVX vector code is introduced; all changes are plumbing only.

Benchmark (Q4_0, SM8850 HTP, NDEV=1, -t 4, tg50 x5):

Before: 18.61 +/- 0.17 tok/s  (224 CPU splits/graph)
After:  20.24 +/- 0.28 tok/s  (0 CPU splits for CLAMP)
Gain:   +10.4%

## Additional information

Same fix benefits any model using `GGML_OP_CLAMP` on the HTP backend (e.g. Gemma 2/3 attention logit softcapping). Changes touch `ggml-hexagon.cpp` (op remap + `device_supports_op` routing), `htp-ops.h` (new `HTP_OP_CLAMP` enum), `htp/main.c` (dispatch), and `htp/unary-ops.c` (row-block and tiled clamp implementations, following the existing `SCALE`/`SQRT` pattern).

## Requirements

<!-- IMPORTANT: Please do NOT delete this section, otherwise your PR may be rejected -->

- I have read and agree with the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md)
- AI usage disclosure: <!-- mention: YES / NO - if yes, describe how AI was used -->

[ggml-gh-bot]

Hi @cqderek, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[cqderek]

@max-krasnyansky could you please review this? This PR adds HTP backend support for GGML_OP_CLAMP.

[max-krasnyansky]

@max-krasnyansky could you please review this? This PR adds HTP backend support for GGML_OP_CLAMP.

Please update the PR description (AI usage disclosure, etc).
Overall looks good.

[max-krasnyansky]

@cqderek I realized that we already had this queued up internally #25934
I'm going to close this one. The implementation is basically the same since we already had hvx_clamp in the library.