Report IsComplete() true for a division-by-zero $size_in_bytes

[AaronWebster]

Part 3 of 3 of the integer division/modulus series.

Stacked on #274 (base branch: emboss/divmod-undefined). Review/merge #273
and #274 first; the base retargets automatically as the chain merges. The diff
below is only this phase's changes.

A structure whose intrinsic size divides by zero has an undefined size: it can
never be Ok(), so per the design doc
(doc/design_docs/division_and_modulus.md, "IsComplete()") it should report
IsComplete() == true -- there is nothing more to wait for -- to distinguish it
from a structure that is merely a few bytes short.

What this does

To carry that distinction, Maybe<T> gains an Unknowability reason
(kKnown / kUnreadable / kUndefined) explaining why a value is un-Known.
A default-constructed Maybe is kUnreadable (the common absent/short-field
case, so most sites need no change); a // 0 or % 0 at runtime is the sole
source of kUndefined. The expression operators in emboss_arithmetic.h thread
and merge the reason, with operation-dependent precedence:

• arithmetic (MaybeDo/Sum/Product/Maximum/comparisons): kUndefined
  dominates -- an undefined input can never be fixed by more bytes;
• short-circuiting And/Or: kUnreadable dominates -- an unreadable
  operand could still resolve the whole expression with more bytes;
• Choice: propagates the condition's reason, or the taken branch's;
• ///%: a Known() zero divisor is kUndefined, else propagates.

Generated views gain Maybe<bool> IsUndefined() (on both the dynamic and
constant virtual-field views), and IsComplete() returns true when
IntrinsicSizeIn*().IsUndefined(). Net effect for a /0 size: Ok() false,
IsComplete() true, IsUndefined() true; a truncated buffer (defined but
unreadable size) still reports IsComplete() false -- the crux distinction.

The reason is a uint8_t enum kept fully constexpr to avoid disturbing the
size-sensitive Ok() path. All C++ goldens regenerate (the shared Maybe<T>
and virtual-view templates changed); the changes are uniform
IsComplete/IsUndefined plumbing. Adds Maybe reason tests, an
emboss_arithmetic reason-merge matrix, and IsComplete/IsUndefined integration
assertions (including the unreadable-vs-undefined case).

bazel test //... is green under both gcc and clang.

---

PR chain (merge in order): #273 → #274 → #275 (this)