riscv64: Split clobber save/restore SP adjustment for compressed instructions

[rajeshgangam]

Summary

• When the Zca extension is enabled, split the SP adjustment in function prologues/epilogues so callee-saved stores use small SP-relative offsets that fit in c.sdsp's 504-byte encoding range
• All callee-save stores/loads compress from 4-byte sd/ld to 2-byte c.sdsp/c.ldsp, saving ~28 bytes per function with 8 callee-saves
• Without Zca, the original combined SP decrement is preserved — zero behavior change on existing tests
• Adds isa_flags parameter to the ABIMachineSpec::gen_clobber_save and gen_clobber_restore trait methods so implementations can make ISA-specific decisions

Before (main, has_c enabled, 1KB frame + 8 callee-saves):

  addi sp, sp, -0x440           ; 4B  combined (clobber+frame)
  sd s1, 0x438(sp)              ; 4B  offset > 504, can't compress
  sd s2, 0x430(sp)              ; 4B
  ...

After (with Zca):

  c.addi16sp sp, -0x40          ; 2B  clobber only
  c.sdsp s1, 0x38(sp)           ; 2B  offset ≤ 504, compressed
  c.sdsp s2, 0x30(sp)           ; 2B
  ...
  addi sp, sp, -0x400           ; 4B  remaining frame

Fixes #7190.

Test plan

• All 235 riscv64 compile filetests pass
• All 116 aarch64, 181 x64, 101 s390x, 49 pulley filetests pass (no regressions from trait change)
• New c-clobber-save-restore.clif test with 3 cases: 100KB frame, 512B boundary, max callee-saves
• Existing 9 riscv64 test files have zero diff against main (non-Zca path unchanged)

[fitzgen]

Hi, can you confirm that you are following our AI policy and have fully reviewed and understood the code here, and are willing to take responsibility for the code in the way that only a human can? Thanks.