Limits: add a sign-neutral finite direction so boundedness is not lost to unknown

[petlenz]

Follow-up to #357 / PR #437.

limit_result::direction (include/numsim_cas/core/limit_result.h) has no value for "converges to a finite limit whose sign is not provable". The only sound answer for such a case is unknown, which downstream also means "might be infinite" — so boundedness information is lost, not just the sign.

Confirmed losses (adversarial review of #437, probes run against the branch):

Expression	True limit	Reported
x + y as x → +∞, y unassumed	+∞ (y is finite)	unknown
exp(y), y unassumed	finite positive	unknown
abs(y), y·y	finite, ≥ 0	unknown
unassumed symbol y	finite	unknown

Before #357 these returned finite_positive — definite but unsound (y may be negative). #437 traded that fabrication for unknown, which is sound but coarse; the user accepted that trade deliberately, with this issue as the follow-up.

Proposal: add direction::finite (converges, sign unknown) and treat it as finite everywhere finiteness is what matters:

• symbols and sign-unknown finite results report finite instead of unknown;
• combine_add: finite_positive + finite_negative → finite; finite ± finite → finite; pos_infinity + finite → pos_infinity (this is the case that currently degrades);
• combine_mul, apply_pow, apply_log, apply_reciprocal: propagate finite, but never claim a sign from it — 1/finite is finite only when the value is also known nonzero, otherwise unknown;
• any is_finite-style predicate counts finite;
• t2s limit visitor mirrors the scalar rules.

direction is public API, so this is a deliberate enum extension: every switch over it must be revisited (the compiler finds them if no default is present — check limit_algebra.cpp, scalar_limit_visitor.cpp, tensor_to_scalar_limit_visitor.cpp).

Soundness rule stays: a reported direction must be true for every value consistent with the assumptions. finite claims convergence only where convergence is provable — an unassumed symbol is finite because it is a fixed real, not because the sign is unknown.

Tests: the table above, plus the two cases #437 changed (ScalarLimit.OtherVariableFinite, T2sLimit.TensorDepIndependentExpr), which would move from unknown to finite.

Signed-off-by: petlenz peterlenz89.pl@gmail.com