chore: bump PLENA_Compiler to merged main (#57 im2col, #58 causal-mask, #59 FFN roll)

[booth-algo]

Re-pins the PLENA_Compiler submodule from 74ff7a5 (#56) to merged-main ebdba9e, so the simulator picks up three compiler fixes that are already merged to PLENA_Compiler main but not yet referenced here:

• ci: cancel stale workflow runs #57 — V_SHIFT_V im2col enabled for non-64-aligned columns (native-dim vision).
• [WIP] Profiling, board configs, weight streaming + memory-footprint (reference) #58 — correct causal mask for multi-tile decoder attention (seq_len > mlen): the static (mlen, mlen) triangular mask is no longer added to every score tile, fixing the sub-64 decoder allclose failures at seq_len > mlen.
• Fix PREFETCH_M_AMOUNT assertion when MLEN > 64 #59 — roll the unrolled FFN inner-K accumulation into a hardware C_LOOP: ~80% fewer ISA lines on the sub-64 decoder, numerically identical.

Pure submodule pointer change; no other files. The renamed ATEN_OPS_UNROLL harnesses already match (the simulator was re-pinned to #56 in #85).

Verification

Smoke-tested the simulator against the merged compiler (ebdba9e), 1 layer, seq_len=64:

case	config	exercises	result
decoder	native_32x32x4	#58 causal multi-tile + #59 FFN roll	PASS (207,677 ISA lines)
vision-layers	native_64x64x16	#57 im2col	PASS (307,648 ISA lines)

The decoder line count reflects #59's FFN roll, and allclose passes (confirming #58's causal fix is active on the combined main).