Add the monoid-algebra spine (#768)

[Rafael-SOWNet]

The representation four features share, per One structure under several features in AGENTS.md (#773). No behaviour change: nothing references it yet.

Three types

Semiring	Zero, One, Add, Multiply, IsIdempotent — instances Field and Boolean
IBasisOps<TBasis>	Identity, Combine, Meet, TryDivide
SparseTerms<TBasis>	immutable; Add, Scale, Multiply, FactorOutCommon

IsIdempotent is why Semiring is a type rather than a pair of delegates. a or a = a holds in the boolean semiring and fails in the complex one, and that single difference is what separates covering from superposition. Anything reading the flag is asking whether it may cover the same basis element twice.

Meet is why this is one abstraction rather than two. Dividing out the meet of the support is simultaneously "take the common monomial out of a polynomial" and "detect that a quantum state is separable":

x^2*y + x^2     meet of (2,1),(2,0) is (2,0)   leaving  y + 1
|001> + |011>   meet of 001,011     is 0-1     leaving  |0> + |1>

What is deliberately absent

Cover selection, factorisation into irreducibles, series truncation. Each belongs to exactly one feature, and a shared engine that swallowed them would be wrong in a way that still type-checks. That list is in AGENTS.md too, so it survives someone finding this class inviting later.

Two things found by building it

• TBasis must compare by value, now documented on the type parameter. Terms are keyed on it, so a basis with reference equality leaves every element distinct and quietly stops collecting like terms — no error anywhere, just a wrong answer. Writing the tests is what surfaced it.
• netstandard2.0 has no Deconstruct on KeyValuePair. The net7.0 build said nothing. Third such catch today, after ToHashSet and BitOperations.

Unverified, and why that is not neglect

Every type here is internal, and InternalsVisibleTo("UnitTests") cannot simply be uncommented — I tried:

error CS1726: Friend assembly reference 'UnitTests' is invalid.
Strong-name signed assemblies must specify a public key in their InternalsVisibleTo declarations.

SignAssembly is on with key.snk, so a friend reference needs the test assembly signed with the same key and its public key named in the attribute. That is a change to the build and to an assembly's identity — the maintainer's call, not mine, so I reverted it.

The tests are written and waiting: two IBasisOps implementations over one value-typed basis — exponent vectors (meet = componentwise minimum) and kets (meet = agree-or-free) — put through the same FactorOutCommon, which is the one claim this whole design rests on. They land the moment the access question is settled.

Nothing references the spine, so the suite is unchanged at 5359 passing. That says the build is sound and says nothing about whether these types are correct.

Reviewer's note: if you would rather not sign the test assembly, the alternative is to let the quantum module be the spine's first test, through its public surface. Merging this as-is is reasonable because it is inert — but it should not be read as verified.