Fix ex_bits=0 slot in excode IP dispatch table

[greenbinjack]

resolve_excode_ip_table() fills index 0 (zero extra bits) with a duplicate
of ip16_fxu1_avx{512,2}, the 1-bit kernel, instead of a function that
contributes zero to the inner product. Since select_excode_ipfunc(ex_bits)
indexes this table directly by ex_bits, any code path with zero extra bits
ends up calling the 1-bit kernel on nonexistent extra-code data, corrupting
the inner product.

This adds a small ip_fxu0 stub that always returns 0.0F and puts it at
index 0 for both the AVX-512 and AVX2 tables, and adds a regression test
(Select_IP_Func.zero_ex_bits_contributes_nothing) that fails on the old
code and passes with the fix.

Tested: full CMake build + ctest suite (32/32 passing), formatting clean
against .clang-format.