hexagon: support for IQ4_NL and MXFP4

[njsyw1997]

Overview

This PR adds support for IQ4_NL and MXFP4 on Hexagon HMX, and fixes a memory address issue that occurs when MXFP4 and Q4_0/IQ4_NL matmul operations are executed consecutively.

Additional information

Due to the lack of a WoA laptop and limited RAM on the test device, I couldn’t run GPT-OSS-20B directly. Instead, I used a Qwen3 variant to validate MXFP4.

The error happens when SKIP_QUANTIZE is enabled. The operator attempts to reuse previously quantized activations. However, the activation memory offset is computed based on the current op’s weight size, not previous op weight size. Now we just record the correct offset for next layer's use.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes. Used for adding tests, logs and creating scripts to filter logs.

[njsyw1997]

By the way, what's the correct way of format code? I tried .clang-format but it does not work.

[CISC]

By the way, what's the correct way of format code? I tried .clang-format but it does not work.

Use git clang-format (after installing the appropriate release of git-clang-format).

[max-krasnyansky]

Looks good.
Only one nit. Misaligned pragma (line 127). clang-format always gets that wrong.

btw Back when I first added MXFP4 support I also just quantized a smaller model (llama3.2-3B at the time) myself to test that end-to-end.

[max-krasnyansky]

@lhez when you get the chance please review/ack.

I tested everything on my devices and looks good.
We're now passing IQ4_NL MUL_MAT tests with Hexagon 👍

$ HB=0 D=HTP0 ./scripts/snapdragon/adb/run-tool.sh test-backend-ops -b HTP0 -o MUL_MAT -p type_a=iq4_nl 
...
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=1,k=32,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  MUL_MAT(type_a=iq4_nl,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): OK
  12/12 tests passed