perf(arrow/ipc): slice already-zero-based offset buffers

[fallintoplace]

Rationale for this change

IPC currently allocates a new offsets buffer whenever a variable-length array has unused trailing offsets. For a prefix slice, the first logical offset is already zero, so the loop only copies every offset and subtracts zero.

Arrow C++ handles this case by slicing the existing offsets buffer.

What changes are included in this PR?

• Read the first logical offset instead of treating every sliced array as needing rebasing.
• Rebase only when that offset is non-zero.
• Use memory.SliceBuffer when the required offsets are already zero-based.
• Add coverage for int32 and int64 offsets, middle slices, and slices with a non-zero array offset whose first logical value offset is zero.

The benchmark uses a 524,288-value slice from a 1,048,576-value backing array. Medians from 6 runs on an Apple M1 Pro were:

prefix slice	main	this PR
int32 time/op	238 us	40.7 ns
int32 B/op	2,106,000	80
int64 time/op	283 us	40.7 ns
int64 B/op	4,203,153	80
allocs/op	8	1

Unsliced buffers remain allocation-free. Middle slices that need rebasing are unchanged.

Are these changes tested?

• go test ./arrow/ipc -count=1
• go test ./arrow/... -count=1
• go test ./arrow/ipc -run "^$" -bench "^BenchmarkGetZeroBasedValueOffsets$" -benchmem -count=6

Are there any user-facing changes?

No. The IPC output and public API are unchanged.