fix(riscv): decode segment fault-only-first loads (vlseg*nf*e*ff)

[carlosqwqqwq]

Allow segment fault-only-first loads vlseg<nf>e<eew>ff to decode.
decode_load_fp currently accepts lumop 0b10000 only when nf == 0
(single-register vle<eew>ff) and rejects every nf != 0 segment form as
illegal, although the V extension defines them as legal instructions
(riscv-v-spec vector-common.adoc norm:vector_ls__seg_ff_unit-stride_op).
QEMU executes the canonical encoding (vlseg2e8ff.v v0, (a0) = 0x23050007)
normally; RAX reports Op::Illegal in both the unit decoder and the
rax-isa-oracle CLI.

Validation

• Assembler check: riscv64-linux-gnu-as -march=rv64gcv emits 0x23050007
  for vlseg2e8ff.v v0, (a0), confirming the encoding is legal.
• QEMU -cpu max runs the witness ELF to completion (exit 0); a control ELF
  with 0xffffffff dies with SIGILL (132).
• RAX probe: decode(0x23050007, Xlen::Rv64, &Isa::rv64gc()) returns
  Op::Illegal before the fix, and the decode arm is reachable after the
  fix with the same Op used by the non-fault-only-first segment loads.
• Regression test added covering nf=1..3 with e8/e16/e32/e64 and both
  unit-stride and strided forms.

Closes #288

[19h]

Thanks for the report and patch. I independently validated the segment fault-only-first encodings and implemented the complete behavior in #296: first-element faults remain precise, later faults trim vl, partial segments are preserved, and direct/SMIR/QEMU paths are covered. The rollup includes your co-author credit and asks for your review. I’m closing this PR as superseded by #296.