parglm 0.1.9

• Add biglm, fastglm, glm2, and mgcv to the timing comparison in the parglm.Rmd vignette. And add some timings for fewer observations and fewer coefficients.
• Bump roxygen2 to 8.0.0
• Remove busy-wait polling in qr_parallel::get_stacks_res
• Return eta and mu from parallelglm to skip R-side recomputation
• Sum only upper triangle of per-chunk Fisher information in FAST path
• Drop const from R_F members to enable move semantics
• Hoist NA-zeroing of beta out of per-chunk workers
• Use thread_local index for thread_pool::get_id
• Skip intermediate matrix when reverse-pivoting R
• Cache parglm_supported keys to avoid rebuilding family objects per fit
• Change the default of parglm.control(nthreads) to be parallelly::availableCores(omit = 1L) to make better use of available cores
• Replace linked-list task queue with deque under single mutex
• Fuse X chunk copy and weight scaling into one memory pass
• Support quasibinomial() and quasipoisson() family specification
• Enable full compatibility with the sandwich package for robust standard errors, including adding comments to the parglm helpfile, a new vignette, and tests
• Add fastglm's method = 3L, and parglm's parglm.control(method = "LAPACK") to the benchmarks in the parglm.Rmd vignette
• Allow for two column response for family binomial and quasibinomial
• Fix issue with order in summary.parglm and vcov.parglm
• Warn rather than error when starting values cannot be found
• Add confint.parglm to fix profile and Wald CI compatibility, and to give compatibility with the gtsummary package (e.g., gtsummary::tbl_regression())
• Add tidy_parglm_robust for tidying a parglm model with robust standard errors
• Add gtsummary examples to the second vignette