fix multivalue ff indexing regression

[PSeitz]

fixes multivalue ff regression by avoiding using get_val. Line::train calls repeatedly get_val, but get_val implementation on Column for multivalues is very slow. The fix is to use the iterator instead. Longterm fix should be to remove get_val access in serialization.

Old Code 8e773ade7715210fb18a9a3b870329eed35dd1ad

test fastfield::bench::bench_multi_value_ff_merge_few_segments                                                           ... bench:  46,103,960 ns/iter (+/- 2,066,083)
test fastfield::bench::bench_multi_value_ff_merge_many_segments                                                          ... bench:  83,073,036 ns/iter (+/- 4,373,615)
est fastfield::bench::bench_multi_value_ff_merge_many_segments_log_merge                                                ... bench:  64,178,576 ns/iter (+/- 1,466,700)

Current

running 3 tests
test fastfield::multivalued::bench::bench_multi_value_ff_merge_few_segments                                              ... bench:  57,379,523 ns/iter (+/- 3,220,787)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments                                             ... bench:  90,831,688 ns/iter (+/- 1,445,486)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments_log_merge                                   ... bench: 158,313,264 ns/iter (+/- 28,823,250)

With Fix

running 3 tests
test fastfield::multivalued::bench::bench_multi_value_ff_merge_few_segments                                              ... bench:  57,635,671 ns/iter (+/- 2,707,361)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments                                             ... bench:  91,468,712 ns/iter (+/- 11,393,581)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments_log_merge                                   ... bench:  73,909,138 ns/iter (+/- 15,846,097)

[fulmicoton]

It would be much more efficient to remove the get_val call and use the iterator here.

let step = /* .. */;
let index = (0..n).step_by(step);
let vals = column.step_by(step);
let positions_and_values = index.zip(vals);

[fulmicoton]

(that requires removing Line::estimate and call Line::train_from directly in the codec.)

[PSeitz]

This are only 20 samples, but span the whole Column, so I'm not so sure this is more efficient. It depends on the iterator implementation if it can skip efficiently.

[PSeitz]

The default impl calls next() https://doc.rust-lang.org/src/core/iter/traits/iterator.rs.html#284-289, that's probably much slower for large indexes.