Rewrite GPU kernels across frameworks

Rewrite GPU kernels: candle-core → kt-native substrate

Mechanical candle-core → kt-native GPU kernel substrate swaps where logic must be preserved byte-for-byte (#1082).

• · Kernel subsystem (tm_conv1d, tm_gdn_chunk, etc.)
• · Function list per subsystem
• · Source metal.rs boundary kernel code

• · Candle-free kt-native rewrites (JSON per subsystem)
• · Preserved logic (byte-for-byte kernel calls)

• · Claude agents (1 per kernel subsystem)
• · Metal GPU framework (preserved)
• · kt-native substitutes (new)

• · 6 disjoint subsystem groups (GROUPS array)
• · Source functions readable in metal.rs
• · Multi-phase: Draft

• · Mechanical rewrite only: must preserve kernel logic exactly
• · If logic is 'improved', the substrate swap is broken
• · Storage.Metal → kt equivalents must use identical buffer indices
• · pipeline/encoder/device APIs must match exactly

• · 6 disjoint subsystem groups (no write conflicts)
• · Per-group agent isolation
• · Critical rule: PURE SUBSTRATE SWAP (logic unchanged)
• · 7-step rewrite pattern (signatures, literals, allocation, device, storage, buffer_o_kt, existing methods)