remove Base.memhash global (#59697)

This became unsound to use even though it was preserved "to avoid
breakage" in v1.13, since continuing to use it would give incorrect hash
results, which could result in corrupt dictionaries and unsound
programs.

Since #59691, these broken `hash` methods in the ecosystem can now
simply be deleted as they no longer provide any value.

It is hard to say whether this is technically breaking or not as a
change. It causes packages to go from giving subtly wrong answers (the
worst kind of wrong) to crashing in v1.13, until the offending incorrect
methods are deleted.

n.b. this is expected to break several packages (notably among them,
JuliaInterpreter, which has several tests specifically just for the
existence of this global), which will require some ecosystem updates.
The update should just be to delete the offending method (they is now
redundant, undesirable, and have been giving unsafely buggy answers),
but we need to do some work to identify those places and release new
versions.

Author: vtjnash

NEWS.md

@@ -17,7 +17,9 @@ Language changes
 * `mod(x::AbstractFloat, -Inf)` now returns `x` (as long as `x` is finite), this aligns with C standard and
 is considered a bug fix ([#47102])
 
-  - The `hash` algorithm and its values have changed. Most `hash` specializations will remain correct and require no action. Types that reimplement the core hashing logic independently, such as some third-party string packages do, may require a migration to the new algorithm. ([#57509])
+* The `hash` algorithm and its values have changed for certain types, most notably AbstractString. Any `hash` specializations for equal types to those that changed, such as some third-party string packages, may need to be deleted. ([#57509], [#59691])
+
+* The `hash(::AbstractString)` function is now a zero-copy / zero-cost function, based upon providing a correct implementation of the `codeunit` and `iterate` functions. Third-party string packages should migrate to the new algorithm by deleting their existing overrides of the `hash` function. ([#59691])
 
 * Indexless `getindex` and `setindex!` (i.e. `A[]`) on `ReinterpretArray` now correctly throw a `BoundsError` when there is more than one element. ([#58814])
 

base/hashing.jl

@@ -629,7 +629,3 @@ hash(data::AbstractString, h::UInt) =
     hash_bytes(utf8units(data), UInt64(h), HASH_SECRET) % UInt
 @assume_effects :total hash(data::String, h::UInt) =
     GC.@preserve data hash_bytes(pointer(data), sizeof(data), UInt64(h), HASH_SECRET) % UInt
-
-# no longer used in Base, but a lot of packages access these internals
-const memhash = UInt === UInt64 ? :memhash_seed : :memhash32_seed
-const memhash_seed = UInt === UInt64 ? 0x71e729fd56419c81 : 0x56419c81