Vectorize se_simple() and se_sum() for performance

[wcurrangroome]

Problem

se_simple() and se_sum() in R/calculate_cvs.R are the most-called functions in the MOE pipeline (invoked hundreds of times inside dplyr::across() in calculate_moes()), but both have unnecessary overhead:

se_simple() (line 8)

Uses purrr::map_dbl(moe, ~ .x / 1.645) for what is natively vectorized as moe / 1.645. The map_dbl wrapper adds per-element function-call overhead for no benefit.

se_sum() (lines 16-61)

For each call, this function:

1. Converts list inputs to data frames
2. Pivots twice (pivot_longer + pivot_wider)
3. Groups and splits by observation
4. Maps 4 times (pull, se_simple, square, sum)
5. Collects with map_dbl(sqrt)

The core operation is sqrt(sum(se^2)) row-wise, with a Census Bureau special case: when multiple zero-estimate observations are summed, keep only the largest MOE among them.

A vectorized matrix-based approach should be significantly faster — but needs careful testing against the current implementation for the zero-estimate edge case.

Suggested approach

1. Replace se_simple() body with moe / 1.645
2. Rewrite se_sum() using matrix operations
3. Benchmark before/after on tract-level data for a large state
4. Verify the zero-estimate edge case produces identical results

Context

Identified during code review. For tract-level data across multiple states, calculate_moes() is a significant portion of total compile_acs_data() runtime.