* add matrixmultiply

needed for nip4 compat mode

* fix matrixload sniff for some matrix files (#4372)

matrixload is_a could fail to detect files like this:

```
2 2
0 0
1 1
```

The header parser was not stopping at EOL and thought that this file had scale
0, which it would then reject.

* fix invertlut in some cases (#4373)

If the measurements filled the entire x range, we were not writing the
final value.

For example:

```
2 2
0 0
1 1
```

```
$ vips invertlut linear.mat x2.v
$ vips getpoint x2.v 255 0
-nan
```

With this PR you get 1, as expected.