Matrix SRAM L-Compute for Mamba/KDA recurrence

[Happymic]

This PR implements Matrix SRAM L-Compute for Mamba-2 and KDA recurrence. The compiler places state tiles in PLENA's fixed diagonal SRAM layout and generates L_TILE instructions to process groups of logical rows. Rust executes those instructions, updates the state, and writes the results through the SRAM and HBM data path.

The Compiler → assembler → Rust path has run four consecutive tokens using prepared synthetic BF16 inputs at the model state dimensions. With the phased layout, all outputs and the final state match the reference exactly, with zero bank stalls. A separate integration test also verifies projection writeback into a Matrix SRAM view.

This is a draft for architecture and implementation review; do not merge. The paired Compiler PR #79 is pinned by the submodule. The completed path is the recurrence core; full-checkpoint execution in Rust and RTL resource mapping remain future work.

Reproduce and implementation notes

On Linux with Nix flakes enabled:

git clone --branch review/matrix-lcompute-20260905 https://github.com/AICrossSim/PLENA_Simulator.git
cd PLENA_Simulator
git checkout 44693edaf041de7ddc922a1f50fdc735cbc59068
git submodule update --init --recursive
nix develop --no-write-lock-file --command just test-matrix-lcompute

The ISA specification records the remaining encoding and hardware questions: viewed M_MM_WO reinterprets old immediate bit 17, and L_TILE's FP32 intermediates still need physical resource mapping.