out_s3: build the chunk debug message only when debug is on

[Watson1978]

out_s3 interpolates its two chunk debug messages at the call site, so dump_unique_id_hex, Fluent::Plugin::Buffer::Metadata#to_s and the concatenation all run on every chunk write — including at the default log level, where Fluent::Log#debug returns on its first line without printing anything. Passing a block moves that work behind the level check.

• Measured with fluentd 1.19.3 on Ruby 4.0.6 over 200,000 calls with a real Metadata and unique_id: at log_level info the current form costs 12 objects / 1000 B / 1.417 µs per call, the block form 1 object / 40 B / 0.044 µs. At log_level debug the two are identical (22 objects / 3240 B / ~2.3 µs)
• log.debug { ... } is enough here. log.on_debug { log.debug ... } removes the remaining Proc allocation and reaches 0 B, but 40 bytes does not pay for the extra nesting
• The logged line is unchanged when debug is on, confirmed through the test driver against the real write path
• This fires once per chunk write rather than per record, so the absolute saving is small. It is included because the fix costs nothing and the allocation is pure waste