BUG: mixed-backend ufuncs use inconsistent common-backend policies

QuadPrecDType.common_instance selects the SLEEF backend for mixed SLEEF and long-double descriptors, since long double can widen to binary128.

Existing operations do not consistently follow that policy:

• np.result_type(sleef, longdouble) selects SLEEF.
• Binary ufunc resolvers select long double.
• Comparison resolvers select long double.
• matmul rejects non-SLEEF inputs.

This makes result dtype and precision operation-dependent for the same pair of input descriptors.

After lossless inter-backend casting is available (see #112 ) , supported mixed-backend operations should consistently follow common_instance, preferably through a shared backend-selection helper.

Acceptance criteria:

• result_type, arithmetic, comparisons, and dot-family loops agree on the common backend wherever both backends are supported.
• Both operand orders produce the same result dtype.
• Precision-sensitive tests verify that promotion does not narrow through long double or double.
• Unsupported operations fail explicitly rather than applying a different promotion rule.