Improve speed of ZSTD_compressSequencesAndLiterals using Neon

[arpadpanyik-arm]

This PR improves the generic implementation of ZSTD_get1BlockSummary and adds a Neon implementation of convertSequences_noRepcodes to speed up ZSTD_compressSequencesAndLiterals. Details are in the commit messages. Unit tests included.

Relative performance of ZSTD_get1BlockSummary with ./fullbench -b19 -l5 enwik5 using GCC-13 as baseline:

    Neoverse-V2   before     after    uplift
    GCC-13:      100.000%  290.527%   2.905x
    GCC-14:      100.000%  291.714%   2.917x
    GCC-15:       99.914%  291.495%   2.917x
    Clang-18:    148.072%  264.524%   1.786x
    Clang-19:    148.075%  264.512%   1.786x
    Clang-20:    148.062%  264.490%   1.786x

    Cortex-A720   before     after    uplift
    GCC-13:      100.000%  235.261%   2.352x
    GCC-14:      101.064%  234.903%   2.324x
    GCC-15:      112.977%  218.547%   1.934x
    Clang-18:    127.135%  180.359%   1.418x
    Clang-19:    127.149%  180.297%   1.417x
    Clang-20:    127.154%  180.260%   1.417x

Relative performance of convertSequences_noRepcodes with ./fullbench -b18 -l5 enwik5 using Clang-18 as baseline:

    Neoverse-V2   before     after    uplift
    Clang-18:    100.000%  311.703%   3.117x
    Clang-19:    100.191%  311.714%   3.111x
    Clang-20:    100.181%  311.723%   3.111x
    GCC-13:      107.520%  252.309%   2.346x
    GCC-14:      107.652%  253.158%   2.351x
    GCC-15:      107.674%  253.168%   2.351x

    Cortex-A720   before     after    uplift
    Clang-18:    100.000%  204.512%   2.045x
    Clang-19:    102.825%  204.600%   1.989x
    Clang-20:    102.807%  204.558%   1.989x
    GCC-13:      110.668%  203.594%   1.839x
    GCC-14:      110.684%  203.978%   1.842x
    GCC-15:      102.864%  204.299%   1.986x

Relative performance of ZSTD_compressSequencesAndLiterals with ./fullbench -b17 -l5 enwik5 using GCC-13 as baseline:

    Neoverse-V2   before     after    uplift
    GCC-13:      100.000%  108.730%   1.087x
    GCC-14:      100.428%  109.235%   1.087x
    GCC-15:       98.730%  106.345%   1.077x
    Clang-18:    107.889%  117.064%   1.085x
    Clang-19:    108.088%  117.354%   1.085x
    Clang-20:    108.027%  117.614%   1.088x

    Cortex-A720   before     after    uplift
    GCC-13:      100.000%  108.104%   1.081x
    GCC-14:       99.860%  108.164%   1.083x
    GCC-15:       98.028%  105.675%   1.078x
    Clang-18:    104.380%  113.002%   1.082x
    Clang-19:    105.117%  113.998%   1.084x
    Clang-20:    105.276%  114.038%   1.083x

An additional commit enables optimized SVE2 builds for QEMU, reducing its runtime from ~27 minutes to ~14 minutes.

Co-authored by, Thomas Daubney Thomas.Daubney@arm.com

[Cyan4973]

This looks good to me.
Performance impact is well documented.
There is a relevant test added to the existing framework.
The modified code is rather well concentrated (notably the neon specific variant).
I think the optimization of ZSTD_get1BlockSummary() is pretty clever, on top of being generic which is very welcome,
but also it's not so obvious to the reader, and a short explanation would been welcome, for future reviewers.

[arpadpanyik-arm]

Thanks for the review! You are right, the internals of ZSTD_get1BlockSummary function is not obvious at first sight, so added some relevant clarifications in 8e44004.