fix: correct parity shifts, guard negative shifts, use original mantissa for 2^k path

[puji4810]

• mul_parity_32,64: use 63 - beta_minus_1 (shifting by 64 when beta_minus_1 == 0 is UB; aligns with delta)
• multiple_of_pow2: guard e2 >= 0 before shifting to avoid invalid shift counts from negative-to-unsigned conversion
• dragonbox_main: check the original mantissa before m2 |= mflags; after setting the implicit bit, m2 is never zero and the branch is unreachable