[FLINK-40005][cdc-cli] CliExecutor application mode resolves pipeline definition from local path, remote path or inline content

[Zhile]

What is the purpose of the change

Fix [FLINK-40005]: in application mode, CliExecutor.main failed to start a YAML pipeline with Missing required field "source" in top-level configuration.

args[0] was passed straight to YamlPipelineDefinitionParser.parse(String, ...), which treats its String argument as YAML content. When args[0] is a file path, the path itself is parsed as YAML, producing a document without source/sink and failing validation.

The two native application deployment executors share CliExecutor.main but pass args[0] with different semantics — the root of the problem:

• K8SApplicationDeploymentExecutor sets APPLICATION_ARGS = commandLine.getArgList(), so args[0] is the pipeline definition file path (shipped into the JobManager container, e.g. mounted via a ConfigMap by the Flink Kubernetes Operator).
• YarnApplicationDeploymentExecutor reads the file on the client side and sets APPLICATION_ARGS to the pipeline definition content.

So a fix that only reads args[0] as a path would fix K8S but regress YARN (whose args[0] is already content, not a path).

Regression introduced by #3643 (FLINK-35360, Support Yarn application mode for yaml job).

Brief change log

• CliExecutor.main now resolves args[0] via resolvePipelineDef(...), in three cases:
  1. Explicit-scheme path (s3:///hdfs:///oss:///file://) — read through Flink's FileSystem so the matching plugin resolves it.
  2. Bare local file path — read with the local JVM file API (Files.readAllBytes) rather than Flink's FileSystem, whose cluster default may be S3/HDFS and would not resolve a local path shipped next to the JobManager.
  3. Otherwise the value is already the pipeline definition content (YARN) — parsed verbatim.
• Fixes K8S application mode without regressing YARN, and additionally supports placing the pipeline definition file on remote storage (S3/HDFS/OSS).
• Update Kubernetes deployment docs (EN + ZH): the FlinkDeployment example now uses entryClass: org.apache.flink.cdc.cli.CliExecutor without --use-mini-cluster, and the "native application mode is not supported" note is removed.

Verifying this change

• CliExecutorTest covers all three resolution paths (local path / scheme path via file://, same code path as s3:// / inline content) plus the original regression.
• Manually verified on a Flink 1.20 Kubernetes Application-mode deployment (Flink Kubernetes Operator): the pipeline starts and runs; previously it failed with Missing required field "source".

Does this pull request potentially affect one of the following parts

• Dependencies: no
• The public API: no
• The serializers / checkpointing: no

[Zhile]

@MOBIN-F can you help to review? I'm not sure if this applies the original design.