Avoid ldexp call when converting random numbers to double

[GuySten]

Description

prn() converts the 64-bit output of the PCG generator to a double in [0, 1) with ldexp(result, -64). This compiles to an out-of-line library call on every random number, and profiling shows it is a noticeable fraction of run time (scalbn/__scalbn together were ~7% of a heavy water problem with thermal scattering and ~5% of a gamma shielding problem).

This PR replaces it with a multiplication by the exact power of two 0x1p-64. Because the conversion from uint64_t to double is unchanged and scaling by a power of two is exact (the result is either zero or at least 2⁻⁶⁴, so it cannot underflow), the result is bit-for-bit identical to ldexp. I checked this on 2×10⁸ inputs, including edge cases such as 0, 2⁵³±1 and 2⁶⁴−1, with no mismatches. The now-unused <cmath> include is removed.

Random number sequences and simulation results are unchanged; the full unit and regression test suite gives the same results as develop, and tallies were bit-for-bit identical in the timing runs below.

Transport time, single thread, average of three runs:

Problem	develop	this PR	Change
Watt source in a 100 cm heavy water sphere with c_D_in_D2O, 20,000 histories	14.13 s	13.13 s	−7.1%
Cs-137 photon source in a W/Pb cask, 100,000 histories	21.29 s	20.39 s	−4.2%

Checklist

• I have performed a self-review of my own code
• I have run clang-format (version 18) on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (if applicable)
• I have made corresponding changes to the documentation (if applicable)
• I have added tests that prove my fix is effective or that my feature works (if applicable)