Improve performance of STARTING WITH with insensitive collations

[asfernandes]

To process STARTING WITH with insensitive collations, it's first necessary to generate canonical bytes of the matching strings.

If the matching string is much greater than the pattern string, a time is wasted generating unneeded canonical bytes.

It's necessary to only generate canonical bytes for the initial substring with the same length of the pattern string.

In my tests with character set WIN1252 collate WIN_PTBR matching strings of length 60 and pattern string with length 1, I see performance improvement of ~30%.

With character set UTF8 collate UNICODE_CI I see performance improvement of ~50% in the same test.

Test:

execute block
as
    declare p varchar(1) character set win1252 collate win_ptbr = 'x';
    declare s varchar(60) character set win1252 collate win_ptbr = 'x12345678901234567890123456789012345678901234567890123456789';
    declare n integer = 0;
    declare b boolean;
begin
    while (n < 1000000)
    do
    begin
        b = s starting with p;
        n = n + 1;
    end
end!

execute block
as
    declare p varchar(1) character set utf8 collate unicode_ci = 'x';
    declare s varchar(60) character set utf8 collate unicode_ci = 'x12345678901234567890123456789012345678901234567890123456789';
    declare n integer = 0;
    declare b boolean;
begin
    while (n < 1000000)
    do
    begin
        b = s starting with p;
        n = n + 1;
    end
end!

[asfernandes]

Updated performance improvement verified in test after changes in the implementation.

[asfernandes]

After small changes in INTL API description for canonical function and small changes in engine I verified more than 50% improvement with UTF8 and UNICODE_CI in the same test, so I'm changing this issue to also optimize multi-byte character sets.

[pavel-zotov]

Currently implemented only for WINDOWS: package 'psutil' can not be installed on Python 2.7 when it is running on Linux (Debian).