[spark] Refactor PaimonSparkWriter write flow

[huangxiaopingRD]

Purpose

Refactor PaimonSparkWriter to make the Spark write flow easier to follow without changing write semantics.
This change mainly:

• splits bucket-mode handling in write into dedicated private methods
• extracts write context initialization from the main write flow
• extracts the shared per-partition write pattern for bucketed and non-bucketed writes
• keeps the existing write topology and commit merge behavior unchanged

Tests

Not adding new tests.
Reason:

• this change is a readability refactor and does not intend to change behavior
• the affected paths are already covered indirectly by existing Spark write, dynamic bucket, update, delete, and merge test suites

[JingsongLi]

cc @Zouxxyy

[JingsongLi]

The task closures below capture this whole WriteContext via ctx.newWrite() / ctx.bucketColIdx. Because the context also stores driver-side objects (SparkSession, the original DataFrame, and preparedData), those objects become part of every executor closure even though they are only needed while building the plan on the driver. This can make Spark closure serialization fail or drag SparkContext/Dataset state into tasks. Please keep WriteContext limited to serializable task inputs, or extract only the needed scalars/functions before mapPartitions and keep SparkSession/DataFrame out of the captured object.