hex-dma: introduce direct-mapped DMA cache that is better suited for HVX FA mask handling

[max-krasnyansky]

Overview

Full LRU cache is an overkill for HVX FA loops.
This PR introduces simple direct mapped cache and updates FA to use that in the HVX paths.
The original LRU cache has been renamed to dma_cache_fa (fully-assoc).

Address issue identified in #29210

Additional information

Some perf numbers from IQ9.

./scripts/snapdragon/run.py --target  ubuntu:maxk@192.168.1.87 --devices HTP0 -- llama-bench -m ../gguf/Llama-3.2-1B-Instruct-Q4_0.gguf -t 6 -ngl 99 --ubatch-size 1024 -p 0 -d 0,2048,4096,8192 -n 64
+ ssh maxk@192.168.1.87 cd ~/llama.cpp && ulimit -c unlimited && LD_LIBRARY_PATH=./lib ADSP_LIBRARY_PATH=./lib GGML_HEXAGON_DEVICES=HTP0 GGML_HEXAGON_OPPOLL=1 ./bin/llama-bench -m ../gguf/Llama-3.2-1B-Instruct-Q4_0.gguf -t 6 -ngl 99 --ubatch-size 1024 -p 0 -d 0,2048,4096,8192 -n 64 --device HTP0
ggml-hex: Loading driver libcdsprpc.so
ggml-hex: FASTRPC_GET_DOMAINS query failed (0x6c), using static CDSP domains
ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
ggml-hex: Hexagon Arch version v73, DMA64 disabled
ggml-hex: device 0: HTP0 (phys=0, virt=0, domain=cdsp:3)
| model          |       size |  params | ngl | threads | n_ubatch | dev   |         test | master    t/s | this-PR   t/s |
| -------------- | ---------: | ------: | --: | ------: | -------: | ----- | -----------: | ------------: | ------------: |
| llama 1B Q4_0  | 727.75 MiB |  1.24 B |  99 |       6 |     1024 | HTP0  |         tg64 |  48.13 ± 0.06 |  50.86 ± 0.10 | 
| llama 1B Q4_0  | 727.75 MiB |  1.24 B |  99 |       6 |     1024 | HTP0  | tg64 @ d2048 |  29.52 ± 0.03 |  32.23 ± 0.02 | 
| llama 1B Q4_0  | 727.75 MiB |  1.24 B |  99 |       6 |     1024 | HTP0  | tg64 @ d4096 |  22.15 ± 0.02 |  24.43 ± 0.01 | 
| llama 1B Q4_0  | 727.75 MiB |  1.24 B |  99 |       6 |     1024 | HTP0  | tg64 @ d8192 |  14.77 ± 0.00 |  16.49 ± 0.00 | 

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, for the code review

[max-krasnyansky]

@lhez for review and ack