Explicitly mark saved registers with .cfi_restore

Although it works without these (we just refer the unwinder to the red
zone), older versions of libunwind seem to have a bug that cause it to
flakily fail to restore rbx without this. I've attempted to bisect the
problem, but the issue is very flaky and I've failed to find the culprit
four times now, so just give up and work around it. Explicit restores
match what we do in other files.

Hopefully this will clear some issues tha fiat-crypto's CI are running
into.

Change-Id: I6a19679a37cad8e93e6dee554b6a9b3b9b4bbe4a
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/62865
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
Auto-Submit: David Benjamin <davidben@google.com>
Commit-Queue: Adam Langley <agl@google.com>

third_party/fiat/asm/fiat_curve25519_adx_mul.S

@@ -156,12 +156,18 @@ adcx r15, r14
 mov [ r8 + 0x0 ], r15
 mov [ r8 + 0x10 ], rcx
 mov rbx, [ rsp - 0x80 ]
+.cfi_restore rbx
 mov r12, [ rsp - 0x70 ]
+.cfi_restore r12
 mov r13, [ rsp - 0x68 ]
+.cfi_restore r13
 mov r14, [ rsp - 0x60 ]
+.cfi_restore r14
 mov r15, [ rsp - 0x58 ]
+.cfi_restore r15
 
 pop rbp
+.cfi_restore rbp
 .cfi_adjust_cfa_offset -8
 ret
 .cfi_endproc

third_party/fiat/asm/fiat_curve25519_adx_square.S

@@ -124,12 +124,18 @@ adcx r13, rcx
 mov [ rdi + 0x0 ], r13
 mov [ rdi + 0x18 ], r10
 mov rbx, [ rsp - 0x80 ]
+.cfi_restore rbx
 mov r12, [ rsp - 0x70 ]
+.cfi_restore r12
 mov r13, [ rsp - 0x68 ]
+.cfi_restore r13
 mov r14, [ rsp - 0x60 ]
+.cfi_restore r14
 mov r15, [ rsp - 0x58 ]
+.cfi_restore r15
 
 pop rbp
+.cfi_restore rbp
 .cfi_adjust_cfa_offset -8
 ret
 .cfi_endproc