fix(reward): grade x=5, scientific notation, and empty \boxed{} on MATH-500 (#107)

[minion1227]

Fixes #107

Problem

normalize_math_answer / _check_math mark several correct MATH-500 answers wrong when the model writes an equivalent-but-unnormalized form. These are frequent answer shapes, so the grader silently under-counts, biasing both the sep-CMA-ES training reward and reported eval accuracy downward — the same false-negative class as the already-fixed thousands-comma bug (#35).

R.score_text("math500", r"\boxed{x=5}", "5")               # was 0.0 -> now 1.0
R.score_text("math500", r"\boxed{5\times10^{3}}", "5000")  # was 0.0 -> now 1.0
R.score_text("math500", r"so the answer is \boxed{}. It is 42", "42")  # was 0.0 -> now 1.0

Fix (three targeted spots)

• Variable assignment — a boxed x = 5 normalized to x=5 and never matched the bare reference 5. Drop a leading single-variable assignment in normalize_math_answer so the value is graded. Guarded with a (?!=) lookahead so == stays intact, and because </> are not word characters, inequalities x<=5/x>=5 are never touched.
• Scientific notation — after \times/\cdot become *, 5*10^{3} parsed as neither float nor Fraction, and the sympy fallback reads ^ as XOR. Rewrite a*10^{b} to the canonical aeb so the numeric path parses it (5*10^{3} -> 5e3 -> 5000.0).
• Empty \boxed{} — extract_boxed returns "", which _check_math committed as the blank answer instead of falling back to the last number in the text the way a missing box already does. Treat an empty/whitespace box like a missing box.

Tests

Adds tests/test_reward_math_forms.py (pure stdlib, no torch/GPU/network):

• the three answer forms grade correct (incl. negative exponents, multi-digit coefficients, \cdot)
• guards: wrong values, x<=5/x>=5 inequalities, and an empty box with no number in the text still score 0
• unit checks that normalize_math_answer drops only a leading assignment (not ==/<=) and rewrites scientific notation

Full root suite: 206 passed; ruff check clean on the changed lines.