Reduce memory usage in Parquet->Arrow decimal column chunk conversion

[danburkert]

This PR reduces memory usage, both in terms of memory used as well as
allocations, in the Parquet->Arrow conversion of Decimal chunks. There
are two optimizations:

1. Instead of using slice::concat to expand buffers to 16 bytes, a
   stack-allocated 16 byte buffer is used instead. This removes an
   allocation per value.
2. Data is expanded from the encoded Parquet fixed-size binary pages
   into a byte buffer, which is then converted to a buffer of i128s. To
   reduce the size of the intermediate byte buffer, this conversion is
   now done page-by-page.

[danburkert]

For reasons I don't fully understand, optimization #2 was yielding the wrong results, so I've backed it out in the second commit and replaced it with a FromIter call.

[codecov]

Codecov Report

Merging #751 (d000eff) into main (2493f7d) will increase coverage by 0.19%.
The diff coverage is 68.42%.

@@            Coverage Diff             @@
##             main     #751      +/-   ##
==========================================
+ Coverage   70.80%   71.00%   +0.19%     
==========================================
  Files         313      313              
  Lines       16930    16912      -18     
==========================================
+ Hits        11988    12008      +20     
+ Misses       4942     4904      -38     

Impacted Files	Coverage Δ
src/io/parquet/read/mod.rs	39.31% <68.42%> (+0.47%)	⬆️
src/array/fixed_size_binary/iterator.rs	83.33% <0.00%> (-8.34%)	⬇️
src/compute/aggregate/min_max.rs	65.90% <0.00%> (-0.76%)	⬇️
src/compute/nullif.rs	0.00% <0.00%> (ø)
src/compute/comparison/mod.rs	39.53% <0.00%> (ø)
src/io/parquet/write/stream.rs	0.00% <0.00%> (ø)
src/compute/comparison/primitive.rs	100.00% <0.00%> (ø)
src/io/json/read/infer_schema.rs	85.57% <0.00%> (+5.08%)	⬆️
src/compute/utils.rs	95.65% <0.00%> (+17.08%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2493f7d...d000eff. Read the comment docs.

[jorgecarleitao]

Thanks! Clippy missing but otherwise ready to ship

[jorgecarleitao]

Noticed an extra optimization we can do here

[danburkert]

Noticed an extra optimization we can do here

@jorgecarleitao nice, good call. I had to insert a chunks_exact() in order to get that to compile, but I think that was the intention.

[jorgecarleitao]

Woops, we are missing an iterator over the values of a FixedSizeBinary. PR here: #757

[jorgecarleitao]

Thanks again, very clean :)