hexagon: add SOLVE_TRI op

[mengshengwu]

Overview

This PR add solve tri op support for hexagon. Use hvx to accelarate the caculation.

• Tests all passes with test-backend-ops.
• Profile kernels on both CPU and HTP0 backends:

Shape	CPU runs	CPU time / run	Work / run	CPU throughput	HTP0 runs	HTP0 time / run	Work / run	HTP0 throughput
ne_lhs=[64,64,4,4], ne_rhs=[32,64,4,4]	11271	128.73 us	2.13 MFLOP	16.55 GFLOPS	16380	81.78 us	2.13 MFLOP	26.04 GFLOPS
ne_lhs=[128,128,4,2], ne_rhs=[32,128,4,2]	3786	275.89 us	4.23 MFLOP	15.32 GFLOPS	8190	144.31 us	4.23 MFLOP	29.29 GFLOPS
ne_lhs=[64,64,8,32], ne_rhs=[64,64,8,32]	354	4008.46 us	68.16 MFLOP	17.00 GFLOPS	1468	8059.77 us	68.16 MFLOP	8.46 GFLOPS
ne_lhs=[128,128,4,32], ne_rhs=[128,128,4,32]	60	17616.20 us	270.53 MFLOP	15.36 GFLOPS	370	17405.15 us	270.53 MFLOP	15.54 GFLOPS
ne_lhs=[256,256,4,2], ne_rhs=[128,256,4,2]	238	7489.87 us	67.37 MFLOP	8.99 GFLOPS	1485	2717.44 us	67.37 MFLOP	24.79 GFLOPS

Perf summary

• HTP0 is faster on 64x64, rhs=32: 26.04 vs 16.55 GFLOPS.
• HTP0 is faster on 128x128, rhs=32: 29.29 vs 15.32 GFLOPS.
• CPU is faster on 64x64, rhs=64, batch=8x32: 17.00 vs 8.46 GFLOPS.
• HTP0 is slightly faster on 128x128, rhs=128, batch=4x32: 15.54 vs 15.36 GFLOPS.
• HTP0 is faster on 256x256, rhs=128: 24.79 vs 8.99 GFLOPS.

Also I tested this with a full model on device using Qwen3.5-0.8B-Q4_K_M.gguf. Here are representative tensor shape samples from the delta-net chunked path around SOLVE_TRI:

common_debug_cb_eval:        attn_pre_solve-21 = (f32) NEG(HTP0#attn-21#0{64, 64, 1, 16}, }) = {64, 64, 1, 16}
common_debug_cb_eval: dnet_add_ch_attn_solved-21 = (f32) ADD(node_2574{64, 64, 1, 16}, HTP0#node_2571#0{64, 64, 1, 1}}) = {64, 64, 1, 16}

common_debug_cb_eval:       dnet_add_ch_lhs-22 = (f32) ADD(HTP0#attn-22#0{64, 64, 1, 16}, HTP0#node_2714#0{64, 64, 1, 1}}) = {64, 64, 1, 16}
common_debug_cb_eval:        attn_pre_solve-22 = (f32) NEG(HTP0#attn-22#0{64, 64, 1, 16}, }) = {64, 64, 1, 16}
common_debug_cb_eval: dnet_add_ch_attn_solved-22 = (f32) ADD(node_2717{64, 64, 1, 16}, HTP0#node_2714#0{64, 64, 1, 1}}) = {64, 64, 1, 16}

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:
  YES, I use AI to learn the API usage of HVX and the algorithms for solving triangular matrix equations and review my code.

[mengshengwu]

Waitting for DMA optimize

[tboinovski1]

I experimented with adding DMA to the kernel and to work off of VTCM but did not see meaningful improvement on v75 and v81.