JIT: Process all register kills in bulk in LSRA (#101760)

`RefTypeKill` is one of the most common types of ref position because we
create a separate one for every register whenever registers are killed,
and registers are killed with very high frequency (for example, all
callee trashed registers at calls). This can be seen in metrics: in
`libraries_tests.run`, `RefTypeKill` is as common as `RefTypeUse` and
`RefTypeDef` combined on x64, and 3x as common as them combined on arm64
due to having more registers.

The main complication around changing `RefTypeKill` to represent killing
a full set of registers is the fact that we want to be able to easily
look up the next `RefPosition` at which a particular register is killed
or used. Today, this is very simple: all kills (`RefTypeKill`) and uses
(`RefTypeFixedReg`) participate in an intrusive linked list, and finding
the next `RefPosition` is as simple as following this list.

This PR changes LSRA to model the kills in bulk instead of creating an
individual `RefPosition` for every kill. To handle finding the next
fixed reg ref position LSRA is updated to take into account that there
can be `RefPosition`'s corresponding to a register from two different
linked lists: the `RefTypeFixedReg` intrusive linked list, and also a
linked list of all `RefTypeKill` ref positions. The latter linked list
may take some searching (not every kill necessarily kills the register
we are looking for), but the reduction in number of `RefPosition`'s
created more than makes up for this.

No codegen diffs, but significant throughput improvements.

This change reduces memory usage of the JIT by around 12% for arm64,
and by around 7% for x64.

Author: jakobbotsch