Selective decoding of a subset (e.g. columns or row groups) of parquet metadata

**Is your feature request related to a problem or challenge? Please describe what you are trying to do.**

As part of https://github.com/apache/arrow-rs/issues/5853 we are considering ways to improve parquet metadata decoding speed in the context of files with wide schemas (e.g. 10,000 columns)

One common observation is that many queries only require a small subset of the columns, but because of how standard thrift decoders are implemented, they must decode the entire metadata even if only a subset of the columns is needed

Due to the Apache Thrift variable length encoding, the decoder likely still requires scanning the entire metadata, but there is no need to create rust structs for fields that will not be read.

Simply skipping such fields I think would likely result in substantial savings. Some evidence for this is @jhorstmann's prototype that avoids copying structs off the stack results in 2x performance improvements. See https://github.com/apache/arrow-rs/issues/5775#issuecomment-2131307588

Thus we could likely optimize the decoding of metadata for large schemas even more by selectively decoding only the fields needed. This idea also described at a high level here: https://medium.com/pinterest-engineering/improving-data-processing-efficiency-using-partial-deserialization-of-thrift-16bc3a4a38b4

**Describe the solution you'd like**
Implement some sort of projection pushdown when decoding metadata. Perhaps we could add a projection argument to this API https://docs.rs/parquet/latest/parquet/file/footer/fn.decode_metadata.html

**Describe alternatives you've considered**

**Additional context**

Start with the parquet::file::footer::decode_metadata API and read the linked arrow-rs issues 5853 and 5775 for the performance context. Define how a projection for columns or row groups should work, then verify that selective decoding preserves the required metadata while avoiding construction of unused Rust structs.

Written by the indexing model from the issue text.