* hexagon: add support for DSV4_HC_PRE
* hexagon: hsv4-hc-pre plain/gated split kernel and l2 prefetch optimization
* hex-dsv4-hc-pre: double-buffered DMA for prefill
* hex-dsv4-hc-pre: refactor DMA block setup
* hex-dsv4-hc-pre: precompute kernel params
* hex-dsv4-hc-pre: support for n_hc != 4
* hexagon: add support for DSV4_HC_POST
* hex-dsv4-hc-pre: refactor gated PRE prefill DMA to dst writeback after each block
* hex-dsv4-hc-pre: remove in-loop vector spills and redundant device-side checks
* hex-run: add support for test-llama-archs tool
* hex-dsv4: refactoring and optimizations
Remove non-dma paths.
Clean up DMA pipelines and dma_addr_t casts.
Add dims collapsing support.
Tested with test-llama-archs -a qwen4ext and test-backend-ops
---------
Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>