repr: fix panic decoding an empty Avro decimal value

[def-]

twos_complement_be_to_numeric_inner read input[0] to determine the sign before checking that the input was non-empty, so a zero-length two's-complement byte run panicked with an out-of-bounds index.

This is reachable from Avro source ingestion: a decimal logical type backed by bytes (or fixed) whose value is encoded as a zero-length byte array is a valid wire form, and that data arrives from external, possibly hostile, schema registries and producers. A conformant producer encodes zero as [0x00], but a non-conformant one can send empty bytes, crashing the decoder and so the source.

An empty big-endian two's-complement run canonically denotes 0. Guarding the sign check with !input.is_empty() makes the existing logic produce 0 for that case (head == 0, no chunks), with no other change needed. The only caller is the Avro decimal decode path.

Found by the avro_decode_fuzzed_schema cargo-fuzz target.

Tests: adds test_twos_complement_empty_is_zero, covering the wrapper at several scales and the generic inner at both numeric widths.

Release note: Fix a panic in Avro-formatted sources when decoding a decimal column whose value was encoded as a zero-length byte array.

[def-]

This fix is already part of #36984. Really need to get this stuff merged or I'll keep rerunning into the same bugs.