vectorize and use torch_inverse and linalg_solve for perspective_boxes performance

[cregouby]

Improve perspective_boxes performance by 48 % :

Reprex :

library(torch)
library(here)
library(dplyr)
library(bench)
library(ggplot2)
devtools::load_all()
startpoints <- list(c(0,0), c(0,200), c(150,0), c(150,200))
endpoints <- list(c(10,20), c(7,120), c(100,0), c(110,180))
boxes <- torch_randint(low = 1, high = 20, size = c(130, 4))
boxes[3:4, ] <- boxes[1:2, ] + boxes[3:4, ] # make max > min

res <- mark(min_time = 5,
            manual_inverse = perspective_boxes(boxes, startpoints, endpoints),
            torch_inverse = perspective_boxes_2(boxes, startpoints, endpoints))
autoplot(res, type = "beeswarm")