Fix pyobj_to_i64 GMP clamping for huge slice indices

__gmpz_get_si truncates big ints to low bits, returning 0 for 1<<333.
This caused del a[9::1<<333] to fail with "step cannot be zero".

Fix: check __gmpz_fits_slong_p before get_si. If value doesn't fit,
clamp to near-maxsize (positive) or near-minsize (negative) based on
__gmpz_cmp_si sign check.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: jgarzik

src/pyo/slice.asm

@@ -163,14 +163,37 @@ pyobj_to_i64:
     lea rax, [rel none_singleton]
     cmp rdi, rax
     je .is_none
-    ; GMP int: use mpz_get_si
+    ; GMP int: check if it fits in i64, clamp if not
     push rbp
     mov rbp, rsp
+    push rdi                     ; save obj ptr
+    lea rdi, [rdi + PyIntObject.mpz]
+    extern __gmpz_fits_slong_p
+    call __gmpz_fits_slong_p wrt ..plt
+    test eax, eax
+    jz .gmp_clamp               ; doesn't fit → clamp
+    pop rdi                      ; restore obj ptr
     lea rdi, [rdi + PyIntObject.mpz]
     extern __gmpz_get_si
     call __gmpz_get_si wrt ..plt
     leave
     ret
+.gmp_clamp:
+    ; Value too large for i64 — clamp based on sign
+    pop rdi                      ; restore obj ptr
+    lea rdi, [rdi + PyIntObject.mpz]
+    extern __gmpz_cmp_si
+    xor esi, esi               ; compare with 0
+    call __gmpz_cmp_si wrt ..plt
+    test eax, eax
+    js .gmp_clamp_neg
+    mov rax, 0x7FFFFFFFFFFFFFFE  ; positive: clamp to near-maxsize (not sentinel)
+    leave
+    ret
+.gmp_clamp_neg:
+    mov rax, 0x8000000000000001  ; negative: clamp to near-minsize
+    leave
+    ret
 .smallint:
     mov rax, rdi
     ret