[REQ] Optimize scalar fallback for shared tile loads #1239

Tile load/store scalar path uses expensive per-element div/mod and lacks benchmark coverage

Problem

The scalar fallback path in copy_from_global / copy_to_global (tile.h) calls coord_from_linear on every element, which performs repeated integer division and modulo to convert a linear thread index into N-D tile coordinates. This is unnecessarily expensive — particularly for 1D tiles (where no coordinate math is needed) and for N-D tiles (where coordinates can be maintained incrementally across iterations).

Missing benchmark coverage

Existing ASV benchmarks (load_store.py, bounds_check.py) and the example benchmark (benchmark_tile_load_store.py) only use POT (power-of-two) tile sizes on aligned arrays, which always hit the vectorized float4 path. The scalar fallback path has zero ASV coverage, meaning regressions would go undetected. Specific gaps:

• No NPOT tile sizes (which force the scalar path)
• No 1D or 3D tile benchmarks
• No structured dtypes (e.g. vec3)
• No distinction between L2-cached and DRAM-bound regimes
• Example benchmark is hardcoded to 2D / tile=32 / float32 with no CLI configurability

Expected outcome

• Scalar path should avoid per-element div/mod where possible
• 1D tiles should use flat indexing with no coordinate math
• N-D tiles should use incremental coordinate iteration instead of recomputing from scratch each iteration
• ASV benchmarks should cover both scalar and vectorized paths across 1D/2D/3D, multiple dtypes, and memory regimes
• Example benchmark should be configurable for profiling arbitrary tile configurations