Raise PrecompileTools compat floor to 1.2 to fix Downgrade CI

[ChrisRackauckas-Claude]

Problem

This PR fixes two main CI failures for LightweightStats.jl on the 1 channel (now Julia 1.12) / JET 0.11.

1. Downgrade Tests resolution failure

The Downgrade Tests job on main failed to resolve the test environment:

ERROR: LoadError: Unsatisfiable requirements detected for package ExplicitImports [7d51a73a]:
 ├─restricted to versions 1.14.0-1 by project
 └─restricted by compatibility requirements with PrecompileTools [aea7be01]
   to versions: 1.0.0-1.8.0 or uninstalled — no versions left
     └─restricted to versions 1.0.0 by an explicit requirement

Root cause: ExplicitImports 1.13+ requires PrecompileTools = "1.2.0 - 1", but this package declared PrecompileTools = "1" (floor 1.0.0). The downgrade job pins every [compat] dep to its floor, pinning PrecompileTools to 1.0.0, which forces ExplicitImports <= 1.8.0 — unsatisfiable against the declared ExplicitImports = "1.14.0" floor.

Fix: Raise the compat floor to PrecompileTools = "1.2". PrecompileTools 1.2 supports julia >= 1.6, so the package's julia = "1.10" floor and the @compile_workload usage are unaffected.

2. JET static-analysis failures (JET 0.11 / Julia 1.12)

The JET error analysis, JET analysis with different numeric types, and JET analysis with complex numbers testsets failed with 4 reports, all originating from cov(::Matrix) (and cor(::Matrix), which calls it). Each report bottomed out at the same Julia-1.12 LinearAlgebra inference artifact: broadcasting (.*) over two contiguous matrix-column SubArrays forces broadcast's unaliascopy path through copyto_unaliased! in LinearAlgebra transpose.jl, where inference yields a Union{Adjoint{T,Union{}}, Transpose{T,Union{}}} and JET flags an "invalid builtin getfield". This reproduces in a ~7-line standalone function with no package code (a Julia/LinearAlgebra false positive), but is straightforward to avoid in source.

Fix: Rewrite both cov branches to accumulate each covariance entry with an explicit index loop over the centered values instead of broadcasting over column/row views. This sidesteps the unaliascopy path entirely — JET now reports zero issues for both report_call and report_opt — and also removes the per-pair temporary allocation.

Latent bug also fixed: the result matrix was allocated as zeros(float(real(eltype(X))), p, p) — a real matrix even for complex input. Whenever an off-diagonal covariance was genuinely complex, assigning it into the real array threw InexactError. The result is now zeros(float(eltype(X)), ...), matching Statistics.cov. A regression test for complex covariance matrices was added.

Local verification

Julia 1.12.6 (the 1 channel) full suite, with JET 0.11.4:

LightweightStats.jl |  455    455   (0 fail)   # was 447 pass / 4 fail before the JET fix

All report_call / report_opt entry points from jet_tests.jl return zero reports on 1.12. Julia 1.10.11 floor full suite: 422 pass / 1 pre-existing broken (the gated JET @test_skip), no regressions. Complex cov verified against Statistics.cov for a matrix with genuinely-complex off-diagonals (matches; old code threw InexactError).

Downgrade resolve (Julia 1.10.11): PrecompileTools=1.0.0 fails (unsatisfiable), PrecompileTools=1.2.0 resolves and loads cleanly.

Please ignore until reviewed by @ChrisRackauckas.