perf(compression): initialize gzip codec state lazily

[lorenzhs]

Rationale for this change

GZipCodec::Init() initializes compressor state and then immediately initializes decompressor state, which tears the compressor state down again. Decompression-only callers therefore pay for an unrelated zlib workspace allocation and initialization.

What changes are included in this PR?

GZipCodec::Init() now validates the compression level and window size without allocating codec state. The existing lazy paths in Compress(), Decompress(), MaxCompressedLen(), MakeCompressor(), and MakeDecompressor() initialize the state they actually use.

The gzip option tests cover compression level and window-size validation independently, including compression level zero.

Are these changes tested?

The Arrow library builds successfully with the enabled compression backends. Existing codec tests exercise one-shot compression, one-shot decompression, and streaming construction in CI, with focused option coverage for the validation retained by Init().

Are there any user-facing changes?

No API or format behavior changes. Gzip codec state is initialized on first use instead of during codec construction.

[tobias-fire]

Stamping, but feels a bit strange to remove the InitCompressor() call from the Init() function completely. I would prefer a solution where we keep it, e.g. by only creating the decompression codec once per range reader, instead of once per page. Maybe that would be a good idea anyways?

[lorenzhs]

It's totally safe, the flow is essentially

auto status = InitCompressor();
assert(status.ok());
InitDecompressor();

and InitDecompressor de-initialises the compressor that InitCompressor just created, so this fix only leaves us without a decompressor at the end -- but everything that needs it ensures it's there.