[salkinium]

Merges the register maps of all 22001 peripheral instances of all 287 STM32
devices into as few variants as possible, to answer how many different silicon
implementations of a peripheral exist, and publishes them on the homepage with
the Peripheral Variant Explorer.

Each variant is one implementation, described by a core that every instance
has and the optional features that only some have. A feature is every
element that appears in exactly the same instances, since those are enabled and
disabled together, like the clock enable bits of a GPIO port that only the
larger packages bond out.

What makes two register maps different

Neither names nor addresses alone can decide this:

• ST renames registers without changing the layout, most visibly between the
  STM32H7x and STM32H7[RS]x. Comparing names alone calls these different.
• ST reuses layouts for different registers. The STM32F1 GPIO_CRL and the
  GPIO_MODER of every other device both start with two-bit fields at the same
  positions, so comparing addresses alone merges the two GPIO implementations.

So both are compared, and not symmetrically: a name at two locations is a
conflict, a location with two names is a rename. Since unrelated elements
never share a name, two more rules catch what names cannot: differently laid out
bit fields that share bits conflict, and a rename only counts inside the
same register. Without those, the STM32L1, STM32L0, STM32G0 and STM32H503
comparators merge into a single variant whose CSR is a pile of unrelated bit
fields.

Three modes

	located by	answers
binary	register address	can one driver poke both at fixed addresses?
source	register name	does code written against a CMSIS header work?
similar	register name, ignoring documentation-only differences	are these essentially the same?

Neither is strictly coarser: source compatibility merges the two GPIO
implementations, but splits COMP, where several devices have a differently
laid out CSR. The similar mode tolerates e.g. USART_BRR described either
as one bit field or as DIV_MANTISSA with DIV_FRACTION.

1808 distinct register maps become 387 binary, 319 source and 282 similar
variants. GPIO comes out as exactly two implementations plus LPGPIO1, which
reuses GPIO_TypeDef with a one-bit MODER.

Feature names

Features are named by the first source that works: a lookup table downloaded
from the gist, a local model served by LM Studio, or a deterministic heuristic.
CI and the homepage only ever use the table, so no build depends on a model. The
published table names 2623 of 2624 features.

Also in here

• header2svd: the value macros of the non-contiguous CRYP_CR_ALGOMODE
  replaced the bit field itself and were spread over three alternate registers.
  Fixed, which changes CRYP_CR on 27 devices and FLASH_OPTR_nRST_MODE on the
  STM32WB1x, and nothing else of the 1.7 million bit fields.
• Register access from the __IO/__I qualifiers, now shown by both
  explorers. Sparse by nature: 136633 members are __IO, 918 are __I/__IM,
  and __O is never used, so write-only registers cannot be told apart.
• The SVD reader and writer keep the structure type as groupName, plus
  descriptions, alternates, dimensions and enumerated values.
• CI generates the variants and attaches variants-archive.zip to releases.

Release fixes from v0.0.4

Also carries two unrelated CI fixes for failures of the v0.0.4 release:

• Publish to PyPI was rejected with Can't have direct dependency: modm-pinout @ git+..., since PyPI does not accept direct URL dependencies.
  It is no longer declared in the docs extra, and the docs workflow installs
  it separately.
• Build docs failed when two deployments raced each other pushing to
  modm-ext/data.modm.io. They are serialized with a concurrency group, and the
  generated content is rebased onto the remote before pushing, since it replaces
  the repository wholesale.

Known limits

• TIM splits mostly on the option registers that route device signals, so those
  are left out of the comparison entirely.
• LPTIM counts as two variants because CMP was renamed to CCR1 along with
  its only field, so nothing indicates that it is the same register.
• Gemma reads TIM COM as "communication" in 2 of 2624 feature names.