Core: Combine 3 GET requests for parquet reads

[varun-lakhyani]

For small file workloads (compaction, manifests, small data files), footer request latency grows linearly with file count and becomes a scaling bottleneck.

Solution

• Introduce SingleFetchInputFile, an InputFile decorator in org.apache.iceberg.io. When a file's size is at or below read.single-fetch-threshold-bytes, newStream() fetches the entire file in one GET and returns a SingleFetchInputStream backed by the downloaded bytes. All subsequent parquet-mr reads - footer size, footer body, row group data - are served from memory with no additional remote calls.
• Files above the threshold, or when the property is unset (0, the default), pass through unchanged with no behavioral difference from today.
• SingleFetchInputStream implements both SeekableInputStream and RangeReadable, preserving parquet-mr's vectorized read path exactly.

[varun-lakhyani]

Combining S3 GET requests alone gives 60-65% improvement, with further gains possible by parallelising them.