Make Hadoop optional for JSON-only COPY INTO deployments

[fupelaqu]

Follow-up from #183 / story R1FIX.4.

Context

R1FIX.4 removed Hadoop from the runtime path for local JSON / JSON_ARRAY reads and format
detection (LocalPath + openStream, java.nio.file only). It deliberately changed nothing in
the dependency graph, so the artefact size is unchanged.

Measured for AC 4 (both jars assembled at the same 0.20.3-SNAPSHOT coordinate):

Artefact	Before	After	Δ
softclient4es8-cli-…-assembly.jar	222 427 032 B	222 434 125 B	+7 093 B (+0.003 %)

The +7 KB is just the two new LocalPath class files. Δ = 0 by construction: no dependency was
added, removed or re-scoped.

Why the number cannot move today

hadoop-client (core/build.sbt:37) and parquet-avro (core/build.sbt:35) remain compile
dependencies of core because three things still genuinely need them:

• PARQUET — AvroParquetReader / ParquetFileReader consume a Hadoop InputFile
• DELTA_LAKE — DeltaLog.forTable takes a Hadoop Configuration
• every remote scheme (s3a, s3, gs, abfs*, wasb*, hdfs) — Hadoop is the remote FS layer,
  and HadoopConfigurationFactory resolves their credentials

Proposal

Offer a JSON-only deployment in which Hadoop is not on the classpath at all — e.g. move
hadoop-client / parquet-avro to Provided behind a json-only build variant, or split the
Parquet/Delta/remote sources into a separate optional module.

This is what would actually move the ~314 MB JDBC-driver figure quoted in #183. It is a packaging
and module-boundary decision, not a bug fix, hence its own issue.

Notes

• Requires deciding the behaviour when a user asks for PARQUET/DELTA_LAKE/a remote URI in a
  JSON-only build — it should fail with a clear "not available in this distribution" message, not a
  NoClassDefFoundError.
• The JDBC-driver before/after measurement belongs to story R1FIX.7, which re-assembles it against
  released upstreams.