Reject nonfinite Q updates before persistence

[Frankie-Xu]

QValueUpdater.update currently writes NaN/Infinity into memory metadata when a reward, bootstrap value, configuration value, or stored Q/EMA is nonfinite. Finite arithmetic can also overflow before persistence. For example, calling update with reward=float('nan') writes a NaN q_value and reward_ma.

Validate the numerical inputs and computed target, Q, and reward EMA before calling text_mem.update. Validate the optional floor before applying it, so it cannot conceal invalid arithmetic. Invalid values raise ValueError and perform no persistence write; valid scalar updates retain the existing formula and floor behavior.

Validation: python3 -m unittest discover -s tests -p test_q_value_finite.py -q passes 3 tests (24 invalid input/overflow cases plus repeated valid updates). These regressions produced 24 failures against main c1b322c. git diff --check passes. Tests load the actual service module and replace only its MemOS type import and persistence boundary; no external MemOS service or end-to-end suite was run.