Default the reflection nodes to 1:R instead of a nothing sentinel

[pbelmans]

Follow-up to #8, on top of the now-merged #9.

#8 spelled "reflect in every node" as nodes === nothing, which then had to be understood in five places. Defaulting to 1:R deletes the sentinel: the kernel test drops from nodes === nothing || s in nodes to s in nodes, and something(nodes, 1:R) disappears from is_singular and _borel_weil_bott.

Dispatch was the alternative, but both cases run the same algorithm and differ only in a predicate, so it would duplicate the fold loop — and the singleton type that avoids the duplication is Base.OneTo reinvented.

Measured, 2000 E₈ weights: 0.3135–0.3142 ms against 0.3141–0.3284 for the sentinel and 0.3145 before #8. Base.OneTo(R) and 1:R benchmark identically in-process, so 1:R wins for reading. Validation stays free: the compiler drops the check loop on the default, where the induction variable is provably in bounds.

Suite, doctests and benchmark/bench.jl green. Node-argument tests 33 → 40, covering out-of-range ranges and empty ones whose endpoints are out of range (9:1).