Add embed_lazy: lazy embedding of operators into composite bases

[shreyasavadatti]

Closes qojulia/QuantumOptics.jl#523

embed always builds the full matrix immediately. For a 12-site chain that's ~1.6 MB per term just to construct the Hamiltonian, before anything gets applied to a state. And TimeDependentSum has no embed dispatch at all, so there's no clean way to lazily build time-dependent Hamiltonians today.

embed_lazy fixes this - you get a LazyTensor or LazySum back, nothing materialises until you call dense or sparse.

What's added:

• src/embed_lazy.jl - six dispatch paths: single-site AbstractOperator, multi-site vector, LazyTensor re-embedding (preserves factor), LazySum (returns LazySum of LazyTensors), and TimeDependentSum (threads through op.static_op and op.coefficients). All paths check basis compatibility before constructing anything. LazyProduct is explicitly unsupported - it breaks time evolution inside LazyTensor per the existing docs.

• test/test_embed_lazy.jl - 13 testsets, 101/101 passing. Covers dense/sparse ops, multi-site, re-embedding with scaled factors, asymmetric bases, LazySum at all sites, error paths, and an allocation benchmark (<4 kB for N=12 vs >1 MB eager).

• Two lines in src/QuantumOpticsBase.jl, one in test/runtests.jl, embed_lazy entry in docs/src/api.md.

One thing to flag: sx + 0.5*sz materialises to a sparse Operator in this library, not a LazySum. embed_lazy on that returns a LazyTensor (still lazy, still correct). To get a LazySum result, construct one explicitly: LazySum([1.0, 0.5], (sx, sz)). Tests are written accordingly.