hexagon: tiling, tracing and optimizations for unary ops

[aparmp-quic]

Overview

This PR extends the "VTCM layout" concept already used for the big Ops (MUL_MAT, FA) (#25425) to pointwise unary/elementwise ops.

• VTCM layout: host CPU selects the layout used to partition VTCM; the NPU uses the same functions to do the actual partitioning, keeping the two in sync and fixing a VTCM overflow on wide rows.
• Kernel params computed on the host: row tiling, thread assignment, and other kernel params are now precomputed on the host and passed to the kernel instead of being derived on-device.
• Specialized loops/functions: per-thread worker functions are specialized to help the compiler generate tighter code for the hot path.
• Removed code duplication: shared HVX kernel logic factored out into hvx-norm.h.
• Fastdivs: runtime integer division replaced with fastdiv in the unary kernels.
• Tracing instrumentation: added for profiling unary ops.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, Antigravity assist for refactoring, profile analysis and drafting

[max-krasnyansky]

@lhez this should be good to go. second review/ACK please when you get a minute.