Support configurable power-of-2 scaling ('oas' vs 'ceil') for floating-point microscaling in QWIX

[copybara-service]

Support configurable power-of-2 scaling ('oas' vs 'ceil') for floating-point microscaling in QWIX

• Support scale_method in qarray.compute_scale_zero_point and HowToQuantize ('default', 'oas', 'ceil') for microscaled floating-point formats ('mxfp8', 'mxfp4').
• Enforce exact Ceil power-of-2 scaling unconditionally for microscaled integer formats ('mxint8', 'mxint4'), disallowing OAS for integers where uniform lattice steps cannot shift without loss of precision.
• Use efficient bit manipulation (mantissa mask add + bitwise and) for 2 ** ceil(log2(scale)) without transcendentals.
• Add scale_method fields to DotGeneralQtConfig for forward and backward passes.
• Add unit tests verifying OAS and ceil scale behaviors for floating-point formats, exact Ceil scaling for integer formats, and INT8 emulation using INT4 passes.