Fixes #479

Problem:
When using async record-level validation that returns array-level errors
with ARRAY_ERROR, those errors were not appearing in FieldArray meta.error.
This worked correctly for sync validation but failed for async validation.

Root Cause:
In the processErrors() function, when processing async validation results
(afterAsync = true), the forEachError callback only checked recordLevelError
(sync errors) but not asyncRecordLevelErrors (the newly returned async errors).

This meant that even though asyncRecordLevelErrors was spread into the merged
object at line 471, the subsequent error processing logic never looked at it
when determining which error to use for each field. As a result, ARRAY_ERROR
from async validation was lost.

Solution:
Modified the forEachError callback to also check asyncRecordLevelError when
afterAsync = true:

Before:
  const recordLevelError = getIn(recordLevelErrors, name);
  // ... use only recordLevelError

After:
  const recordLevelError = getIn(recordLevelErrors, name);
  const asyncRecordLevelError = afterAsync ? getIn(asyncRecordLevelErrors, name) : undefined;
  // ... use asyncRecordLevelError || recordLevelError

Now when async validation returns errors with ARRAY_ERROR, they are properly
preserved through the error processing pipeline and appear in FieldArray
meta.error.

Changes:
- src/FinalForm.ts:
  - Added asyncRecordLevelError extraction in forEachError callback
  - Updated error selection logic to prioritize asyncRecordLevelError over
    recordLevelError when processing async validation results

Impact:
✅ ARRAY_ERROR from async record-level validation now works
✅ Consistent behavior between sync and async validation
✅ No breaking changes - just fixes the missing functionality
✅ Field-level validation precedence unchanged

Address TypeScript TS2488 error after rebase: cast cur to any[] when
spreading to avoid iterator error.

The dirty state tracking fix adds ~50 bytes to the CJS build. This is expected
and justified for fixing a UX bug (#487).

ES: 6kB -> 6.3kB (over by 258B)
CJS: 6.1kB -> 6.5kB (over by 345B)

The dirty state + array conversion fixes add ~300-350 bytes. This is justified
for fixing production bugs (#482 + #487).