Use windowed shift-and-mask digit extraction in GetDigitAtIndexForBase

[pascoec]

The scalar primitive recomputed a floating-point log2 on every call and
assembled each digit one bit at a time; the NativeVector overload then
invoked it per coefficient through the shared library, so nothing hoisted
or vectorized. Extract the digit as a single shift-and-mask window
instead, with the shift and mask hoisted across the vector so the loop
auto-vectorizes. Digits are unchanged everywhere the old code was
defined; windows past the word width now return 0 instead of reading
past it (previously undefined for partial top windows, e.g. a 60-bit
modulus with 13-bit digits).

BV key switching consumes this on every relinearization and rotation.
On Ice Lake (Xeon 8360Y) at 1 thread across {clang-18, gcc-14} x
{native-opt on, off}: the primitive is 46-150x faster,
NativePoly::BaseDecompose 21-150x, DCRTPoly::CRTDecompose 1.45-2.7x,
and BGV BV-mode EvalMult+relin / EvalRotate 1.22-1.90x, with no
regressions; the gains hold at 36 threads under clang (1.19-1.83x).
Verified digit-for-digit against the previous implementation and by
exact reconstruction; all core and pke unit tests pass.