[OPENCL] Implement PFAC kernel: one thread per text offset

[hubertwojcik]

Goal

Implement the core PFAC OpenCL kernel where each work-item starts from a unique text offset and independently traverses the automaton until a dead state or match is found.

Tasks

• Write kernels/pfac.cl: each work-item i starts at text[i] and walks the automaton
• Write match output to a result buffer: matches[i] = pattern_id or -1
• Choose global work size = text length; tune local work size for M4 Pro
• Collect and compact match results on the host side
• Verify correctness: results match sequential baseline on all test inputs

Acceptance Criteria

• PFAC kernel produces correct match output on all test inputs
• Kernel compiles without warnings under Apple's OpenCL compiler
• End-to-end GPU execution (transfer + kernel + readback) measured and logged