modexp: GCC 15 and 16 keep the addmul() product in a stack slot #1801

With GCC 15, which CodSpeed uses since #1798, modexp<expmod_execute_evmone>[mod_len:32/mod_tz:254/exp_bits:256] is 16% slower than with GCC 14 (0.84x), and master runs 9% more instructions on it (135k vs 124k per call).

The difference is in the addmul() loop inlined into modexp_pow2(). GCC 15 and 16 keep the uint128 product of umul() in a stack slot, storing and reloading it in every iteration (16 instructions per word instead of 12):

mul    %r12
mov    %rax,(%rsp)
mov    %rdx,0x8(%rsp)
mov    (%rsp),%rdx
add    (%rdi,%rcx,8),%rdx
...
mov    0x8(%rsp),%rsi

GCC 14 keeps the product in registers on master, but compiles the same loop into this code after unrelated changes, e.g. the INTX_REQUIRE hints of #1794. A fix could keep the product in registers explicitly in addmul().