feat(graph): detect negative-weight edges in PaymentGraphChecker

[woahwhattheheck]

Closes #773.

PaymentGraphChecker validated reachability but had no notion of edge
weight, so a negative-weight hop passed through unnoticed. A greedy traversal
treats a negative hop as if it adds funds, which makes a route containing
one repeatable for unbounded extraction.

Changes

src/graph/PaymentGraphChecker.ts:

• checkGraph(graph) — returns { valid, issues }, reporting every edge whose
  weight is negative. Each issue carries code, from, to, weight, and a
  message naming the offending edge as from -> to.
• New exported types: PaymentGraphEdge, PaymentGraph,
  NegativeEdgeWeightIssue, GraphValidationResult.

Zero-weight edges remain valid — they are legitimate pass-through hops, so the
check is strictly weight < 0, not weight <= 0.

Two details worth calling out:

• number and bigint weights are both handled. Amounts elsewhere in the
  SDK are bigint (Recipient.amount, Invoice.funded), but graph weights
  are natural to express as plain numbers, so checkGraph accepts either and
  compares against the matching zero.
• All offending edges are reported, not just the first. Returning the full
  list means a caller fixing a malformed graph sees every problem in one pass.

checkGraph accepts either { edges } or a bare PaymentGraphEdge[], since
the issue did not pin the shape and both read naturally at a call site.

Validation

• npx vitest run test/paymentGraphChecker.test.ts — 15/15 passing (the 7
  pre-existing cases plus 8 new ones covering: positive weights accepted,
  zero weights accepted, empty graph, a single negative edge, the message
  naming source and target, negative bigint, multiple offenders reported in
  edge order, and the bare-array form).
• npx tsc --noEmit reports 210 errors on this branch and 210 on unmodified
  main — identical, so this change adds none, and zero are in
  src/graph/PaymentGraphChecker.ts. That baseline is pre-existing (missing
  exported members in types.js/priceOracle.js/soroban/footprint.js and
  noUncheckedIndexedAccess nullability) and is not a prerequisite here; I
  mention it only so the red lint isn't read as coming from this PR.