Add net11 ordinal casing APIs - #585

Adds the ordinal casing family from net11:

• char.ToLowerOrdinal(char) and char.ToUpperOrdinal(char)
• string.ToLowerOrdinal() and string.ToUpperOrdinal()
• ReadOnlySpan<char>.ToLowerOrdinal(Span<char>) and ToUpperOrdinal
• Rune.ToLowerOrdinal(Rune), Rune.ToUpperOrdinal(Rune) and Rune.Equals(Rune, StringComparison)

API count 1038 to 1047.

Ordinal casing is not invariant casing

Worth stating up front, because the obvious implementation is wrong. Ordinal casing declines any mapping that would move a character out of its ordinal upper-casing class, so that it stays consistent with OrdinalIgnoreCase. Six BMP code points differ from invariant casing:

The implementation follows the algorithm the runtime itself uses when the ICU ordinal table is not available (TextInfo.PreserveOrdinalLowerCasingClass): invariant casing, with the lower case mapping discarded when it would change the character's ordinal upper case form, plus the long s exception on the upper case side.

Verified against net11 over the whole BMP: it agrees on every code point, for both directions.

Two further behaviours were pinned against the runtime rather than assumed:

• Casing applies per scalar, not per char, so surrogate pairs are mapped. "\U00010400".ToLowerOrdinal() returns U+10428.
• The span overloads return the number of chars written, return -1 when the destination is too small, and throw InvalidOperationException when the buffers overlap. The overlap check runs first.

Divergence

The mapping is derived from the running framework's invariant casing table, so it follows that framework's Unicode version rather than the one net11 is built against. Scripts added in a newer Unicode version than the host will not be cased. Recorded as a //Note: on each member.

Also fixed

char.Equals(char, StringComparison) was already polyfilled and used ToUpperInvariant for OrdinalIgnoreCase, so it reported the long s as equal to s, where net11 reports them as not equal. Now routed through ordinal casing.

Scope

Rune does not exist before netcoreapp3.0, and is not part of netstandard2.1, so those three members are gated to NETCOREAPP3_0_OR_GREATER. The char, string and span members are available on every target framework.

TextInfo.ToLower(Rune) and ToUpper(Rune) are left out. They are culture aware casing rather than ordinal, so they do not belong with this set.

Verification

• Solution builds clean in Release. Consume builds clean in Debug across all 22 target frameworks, which covers the non-FeatureMemory and pre-Rune paths.
• Tests green on net11.0 (1620), net462 (1597), net8.0 (1617) and net10.0 (1620), plus PublicTests, EmbeddedTests, UnsafeTests, NoRefsTests and NoExtrasTests.
• The same assertions run against the BCL on net11 and against the polyfill on earlier target frameworks.