【wip】[common][core][s3] Add provider-neutral FileIO contract suite

[sundapeng]

Summary

This PR adds a provider-neutral, opt-in contract suite for FileIO and applies it to local, Hadoop-local, HDFS, and S3 providers. It follows the layered “Scheme B” design so provider behavior, default-method behavior, API ownership, returned-object behavior, and consumer assumptions are tested independently.

The public FileIO interface, its Javadocs, and the existing FileIOBehaviorTestBase are unchanged.

Changes

• Add a provider-core contract for streams, status and listing, deletion, rename, copy, UTF-8/atomic helpers, and two-phase output lifecycle.
• Add request-aware default-method tests backed by a deterministic recording implementation. These verify delegation, overwrite propagation, resource cleanup, lazy listing, atomic temporary-file cleanup, and retry behavior without adding metadata preflights.
• Classify all 37 public methods declared by FileIO into exactly one owner: provider core, default methods, provider lifecycle, factories, or optional capabilities. Factory selection/discovery and unsupported-capability behavior are covered separately.
• Cover returned-object contracts, including FileStatus defaults and SeekableInputStream.wrap forwarding and unsupported positioning.
• Add the test-only StrictContractFileIO consumer guard. It rejects non-portable caller assumptions such as listing non-directories or renaming to an existing destination, and is exercised by catalog, commit/abort, branch, and tag workflows.
• Extend the existing local, Hadoop-local, HDFS, and MinIO-backed S3 behavior suites from the new opt-in contract base.

Production fixes exposed by the contract

• LocalFileIO.copyFile(..., false) now reports an existing destination instead of silently succeeding.
• Local Hadoop tryToWriteAtomic now preserves an existing target and returns false; object-store schemes do not gain this preflight.
• Renaming two-phase output now discards only its own staged file, preserves concurrently published targets, and performs overwrite fallback only while its staged file still exists.
• FormatTableCommit rolls back writer-owned UUID targets even when a remote commit publishes and then throws.
• S3 multipart upload now obtains its WriteOperationHelper from the owning S3AFileSystem, including Hadoop's required multipart callbacks and filesystem context.

Contract boundaries

The suite deliberately leaves provider-specific edge cases unspecified, including ambiguous rename shapes such as a missing source, an existing destination, identical paths, or an invalid destination parent. It also allows missing-file input failure either while opening or on the first read, does not define listing order or pre-close visibility, and does not prescribe exact timestamp precision.

Request-aware tests constrain logical FileIO delegation where it affects behavior, but do not freeze backend-specific HTTP HEAD/LIST request counts or ordering. No provider profile, vendor flag, or new production capability API is introduced.

Testing

• Passed 243 tests in paimon-common.
• Passed 119 tests in paimon-core.
• Passed 6 focused S3 tests.
• Compiled all 17 filesystem modules in the 22-module Maven reactor.
• Passed Spotless and diff checks for all changed modules.
• The full MinIO-backed S3FileIOTest suite was not rerun locally because Docker was unavailable; the pushed PR runs it in CI.