optimize f_pow when base is int

[GGrush-Doudizhu]

Optimize f_pow when the base is int and the exponent is EUDVariable.

• Add fast paths for bases 0 and 1.
• Factor powers of two from even constant bases:
  a**b = odd_part**b << (trailing_zeros * b).
• Preserve the existing generic _pow path for unsupported cases.

[armoha]

Unlike _pow, this generates multiple triggers on every call sites.

(1) _pow(0, EUDVariable) case is uncommon and not that slow to avoid using it.

(2) pow(int, EUDVariable) branch has unnecessary copies.
When a is not a power of 2, _pow(odd_part, b) is called anyway. The cost of _pow(a, b) depends on the 1-bits in b so _pow(odd_part, b) has same cost as original _pow(a, b).

(3) When a is a power of 2, the _constmul (full 32 bits multiplication) + f_bitlshift replacement isn't clearly better than _pow. f_bitlshift(1, EUDVariable) is not specialized and it rather fallbacks to f_bitlshift(EUDVariable, EUDVariable)` in current implementation.