[Jahongir-Qurbonov]

vec pop and remove shift the remaining items with a loop like items[i] = items[i + 1]. GCC turns this loop into a memmove call (as a tail call), but clang 21 doesn't and emits a slower element-by-element loop. This PR calls memmove directly in vec_template.c, vec_t.c and vec_nested.c. The reference counting logic is unchanged.

Benchmark: pop(v, 0) repeated n/2 times on a copy of a vec with n items, librt built at -O3, CPython 3.13, Linux x86_64. Each number is the best of 15 runs, in ns per pop. pop(v) (no shift) takes ~75-90 ns in every build.

	clang before	clang after	gcc before	gcc after
vec[i64], n=16	89-94	94-96	96-99	95-96
vec[i64], n=1000	173-178	130-132	128-131	126-132
vec[i64], n=10000	1187-1255	786-795	781-792	782-790
vec[str], n=1000	167-177	142-144	135-142	135-139
vec[str], n=10000	972-1040	795-799	795-801	790-805

GCC builds don't change, and clang builds (such as macOS wheels) now match them.

Testing: the existing vec run tests cover pop with an index and remove for all vec kinds, including the shared buffer and decref ordering cases. They pass with both gcc and CC=clang.