fix(provider): tweak reasoning option driven max budget variants

[rekram1-node]

Summary

• always derive high and max variants for token-budget reasoning controls
• use half of the effective allocation for high and allocation minus one for max
• default missing models.dev maximums to a 32,000-token allocation while respecting model output limits

Testing

• bun test test/provider/transform.test.ts --test-name-pattern reasoningVariants
• bun typecheck