Remove unused allow attributes, that were required for false positive by
clippy that has been fixed in new rust versions.

Previously the necessary validations were missing, which could lead to
crashes (#30) or accepting incorrect code.
Now branches, disciplines and branch access are properly validated with
helpful diagnostics; fixes #30.

Previously analysis always returned true because the appropriate
information was not available with OSDI.
With the latest additions to the interface the ANALYSIS_<..> flags are
set by the simulator to communicate the return value of this function.
This commit implements analysis so that is returns true for the analysis
modes listed in the language standard when the appropriate flags are
set.

Also contains various fixes for other bugs (fixes #31).

Add missing change log entry for c02c018d

The Verilog-AMS LRM specifies fmt literals that are slightly different to
those support by the libc implementation. This commit implements a
parser for these supported format arguments to ensure any used format
literals is valid and all arguments have the correct type. During hir
lowering and codegen the format literals are converted to C compatible
format literals. Some extra functions are used for this where conversion
is non trivial.

Previously using $limit would cause a panic during type check.
Now OpenVAF tests that $limit is used correctly.
This is only possible because it assumes that built-in function only have
real arguments.

For now $limit(x,..) is lower the same as x in hir_lower.
However, all type checking logic is implemented so the foundations
to fully support $limit in OSDI are now available.

A common pattern in compact models is that equations are formulated in
terms of some variables that then depend on multiple biases (unkowns).
Handwritten models often perform better than Verilog-A compiled models
because humans recognize this pattern and only derivative by the
intermediate variable and then apply the chainrule to get all required
derivatives.

This commit implements an SSA graph based algorithm that finds subgraphs
of the SSA where such an optimization can be applied. Finding a perfect
solution is probably an NP-Hard problem. Instead a conservative
heuristic is used.

The reults are very promising. All tested models show instructions
counts decrease of at least 22%. Most models even show an increase above
30%. In some extreme cases like BSIMSOI, the instruction count is even
reduced by 60%.

These instruction reductions are often clustered and therefore decrease
register/cache pressure. Furthermore, these decreases also often happen in
"hot" parts of the model. So overall significant performance increases
are achieved. This change is a big step towards closing the gap between
handwritten C code and Verilog-A compiled models.

The changes in the previous commits changed $temperature so that it was
read from memory instead of being treated as an SSA value during the eval function.

However, the wrong GEP indices ([6] instead of [0, 6]) were used.
This commit fixes that and introduces a function that automatically
generates these struct_gep indices to avoid similar bugs in the future.

Previously, $bound_step was always called during the eval function.
However, most $bound_step calls are constant and can be fully handled
during device setup. This save a bit of performance as it repeated
stores during the newton analysis.

Most models do actually not use $bound_step at all so in those cases
OpenVAF now also does not allocate any space for the time step in the
instance data. This can save some memory for very large netlists.

Previously OpenVAF decided whether a separate unknown is required for a
call to ddt during hir lowering. To that end OpenVAF tracked
whether values are op dependent during HIt lowering.
This algorithm didn't work well, because some values are
still sentinel values that are only replaced by their true value later.

This commit removes most of that logic. Now all ddt calls generate an
implicit unkown. That unkown is then removed during residual generation
if that is possible. During residual construction, the MIR is fully
constructed and which values are operating point dependent is also known
(determined by an robust analysis).

In practice that means that OpenVAF will no longer emit unneeded implicit
unknowns for ddt in some edge cases.